The CISO Cheat Sheet to Microsoft Digital Defense Report 2026
I know we’re all busy, so I’ve pulled the most important data from this year’s Microsoft Digital Defense Report into one short read. The full report is well worth exploring, but this article focuses on the numbers and trends most relevant to security leaders. This year’s edition shows that AI is lowering the time, expertise and cost required to discover and exploit weaknesses, enabling attackers to operate with greater speed, scale and autonomy. The report describes this as AI “changing the physics of cybersecurity”: attack timelines are compressing and agentic systems are beginning to automate more of the attack lifecycle. Trend 1: Time has collapsed The clearest finding is about speed. Vulnerability discovery is increasingly running ahead of remediation, while the time from discovery to weaponisation is shrinking. In the AI era, weaponisation can follow in hours. Vulnerability discovery and weaponisation now often needs only a well-written prompt. The median time from discovery in the wild to weaponisation is below 24 hours. Vulnerabilities were on track for a record estimated 72,000 CVEs in 2026. The report expects a multi-year period in which the number of known but unpatched vulnerabilities rises sharply before eventually declining. The median time from container start to first exploit attempt is 5.3 hours. The ransomware actor Storm-1175 moved from web-facing exploitation to deploying Medusa ransomware in as little as 24 hours. CISA added over 110 CVEs to its Known Exploited Vulnerabilities catalogue between November 2025 and May 2026, most within a week of disclosure. The mismatch with enterprise reality is the uncomfortable part. The report says the disclosure-to-exploitation window has shrunk to single-digit days. Between November 2025 and May 2026, CISA added more than 110 vulnerabilities to its Known Exploited Vulnerabilities catalogue, most within a week of disclosure. In one example, Storm-1175 moved from exploiting a web-facing vulnerability to deploying Medusa ransomware in as little as 24 hours. Microsoft’s Red Team makes an important point: a patch is also a disclosure. Every fix points attackers towards the flaw it closes. As AI-assisted analysis makes reversing patches faster and more reliable, the report argues that patch velocity, not simply patch availability, becomes the control that matters. Trend 2: Attack chains are gaining autonomy The report charts a clear progression: September–November 2025: AI assists in orchestrating a campaign, as in the Anthropic-reported espionage case. April–June 2026: Full-chain autonomy is demonstrated in labs. July 2026: Real-world autonomous attacks of increasing complexity appear. Two frontier models, Mythos and GPT-5.5, completed a 32-step attack chain against an emulated enterprise with no human help. The test environment had no defenders. The UK AI Security Institute reached a similar finding independently. GPT-5.5 was the second model to solve its end-to-end corporate network attack simulation, a task AISI estimates would take a human about 20 hours. Open-weight models lag closed models at orchestrating attacks by about seven months. In the wild, early July 2026 saw the first documented automated ransomware extortion attack, named JADEPUFFER by Sysdig. Microsoft has seen AI-orchestrated intrusions that share elements with it. Initial access consistently came from internet-exposed services running known-vulnerable software. The report is measured on this point. Volumes of AI-orchestrated activity remain low. In most observed real-world operations, complex end-to-end intrusions still retain meaningful human direction. The Red Team's framing is the most useful one: AI doesn't create a new class of attack, it collapses the effort needed to find and exploit existing attack paths. Trend 3: AI has become an attack surface The report's second takeaway is that attackers are already targeting AI itself. The adoption numbers explain why. 88% of enterprises are experimenting with agents, and 82% of leaders plan broader rollouts within 12–18 months. The report cites industry projections of roughly 1.3 billion agents in production by 2028. What attackers are doing to AI systems: Four techniques accounted for roughly 90% of observed attack activity in telemetry from Azure AI workloads: malicious link injection (52%), jailbreaks (16%), tool and agent abuse (12%), and identity and credential abuse (10%). Memory poisoning: An attacker who can place content anywhere an agent reads can often place it where the agent remembers. Email is the dominant pathway the Red Team has observed. Human-in-the-loop isn't a complete control: Red team testing saw approval prompts ignored, or bypassed by a policy stored in poisoned memory. Developer tools turned against their owners: The "s1ngularity" malware hijacked locally installed AI command-line tools to hunt for secrets. It leaked about 2,000 secrets and 20,000 files across 225 victims. AI browsers: In daily use at more than 40% of organisations in Microsoft's dataset, they were the initial infection vector in incidents involving 57 malware families in a May 2026 analysis. Chat histories as a target: Malicious AI-assistant extensions reached nearly 900,000 installs and more than 20,000 enterprise tenants, harvesting AI chat histories that often contain code, architecture and credentials. The identity implications are direct. Each agent needs its own identity, not one borrowed from an employee, and a named human sponsor. The report also treats revocation latency, the time to cut off a compromised agent, as a measured outcome. Trend 4: The front door is still identity AI changes the speed and personalisation of social engineering. It hasn't changed where attacks start. Initial access in Defender Experts data: User execution: 30% Valid accounts: 20% Malicious copy-and-paste: 13% Phishing: 11% Together, user execution and valid accounts account for over half of observed activity. In incident response engagements, phishing as an initial access vector jumped from 7% to 23% year on year. The standout numbers: ClickFix: Defender saw user-pasted malicious commands on more than 1.1 million unique devices between February and early May 2026, roughly an 8x increase. AiTM: URLs accounted for 45% of the identified phishing techniques in the report’s breakdown. Separately, the share of AiTM phishing and token-theft attacks more than doubled during the first six months of 2026, although the report notes that their overall share remained comparatively small. Teams vishing: Weekly confirmed malicious volume increased 502% year on year. Per attempt, 48% of vishing calls kept the victim engaged for more than 20 seconds, and 30% lasted more than one minute. When a call succeeds, attackers can persuade the victim to install a remote-management tool and complete reconnaissance in two minutes or less. Business contact impersonation (BCI): The report introduces this term for conversations that start with a trusted-looking contact. Microsoft detected more than 46 million attacks in 12 months, with a 121% spike in April 2026. The share of first-contact messages that explicitly asked for money or documents fell from 17% to 3%. The first message is deliberately bland, like "Are you at your desk?" Password attacks declined 26% year on year, which the report describes as strong evidence that MFA and passkeys are forcing attackers to pivot. That pivot includes techniques such as fatigue attacks, AiTM phishing and token theft designed to circumvent conventional MFA. Trend 5: Attackers are inheriting trust rather than breaking in The report's headline theme is that risk is increasingly interconnected. Threats travel through the trusted relationships between identities, software supply chains, partners and cloud platforms. The numbers: Supply chain: Major open-source ecosystems saw attacks roughly every ten days, accelerating through Q2 2026. Shai-Hulud 2.0 affected 25,000 repositories. Megalodon backdoored 5,561 GitHub repositories in under six hours. Trusted signing: In May 2026, Microsoft disrupted Fox Tempest, a "malware-signing-as-a-service" operation that obtained genuine certificates for ransomware loaders and infostealers. Microsoft revoked over a thousand fraudulent code-signing certificates. Legitimate access: An unsanctioned remote monitoring tool is now one of the most reliable mid-stage indicators of a human-operated intrusion. Cloud identity: Cloud identity abuse was the most observed technique against critical infrastructure, representing 78% of the technique mix in Microsoft’s data for March to June 2026. Hybrid intrusions: The Storm-2885 case showed an attacker using a privileged cloud identity to close and downgrade security alerts. The report says identity governance and alert-handling workflows deserve Tier 0 treatment. Dwell time: DART data shows a statistically significant increase in dwell time year on year, meaning organisations took longer to detect intrusions. Ransomware shows the same ecosystem pattern. Qilin led leak-site data with 16.4% of victims, up 386%. Defender detections tell a different story, with Akira at 22%. Trend 6: Defence works best when it's connected The report is not only a catalogue of threats. The strongest defensive evidence: Disruption works. Following the March 2026 disruption, Tycoon2FA activity had fallen 95% from its November 2025 peak by June. Its share of CAPTCHA-gated phishing declined from 76% in December 2025 to 12% in June 2026. Correlation is the highest-leverage variable. The report argues that how well defenders share and correlate signals across endpoint, identity, cloud, email and network is one of the highest-leverage strategic variables under their control. AI is also changing defensive operations. The report describes AI-assisted triage, incident summarisation, hunting-query generation and detection engineering as ways to accelerate investigations and allow analysts to concentrate on higher-value work. Automated disruption can compress response time. The report says that when a correlated incident reaches a defined confidence threshold, defensive systems can automatically isolate devices, disable compromised accounts, revoke sessions or quarantine affected cloud resources. It also stresses the importance of appropriate governance and human oversight for high-risk AI actions. The report also recommends changing what leaders measure. It suggests moving from patches deployed to exposure reduced, detection coverage increased and time-to-mitigate compressed. What the MDDR recommends: ten security priorities The report starts its recommendations with the foundations: Secure by Design, Secure by Default and Secure Operations, applied through identity-centric Zero Trust. Microsoft says these keep the priorities below durable as AI adds more identities, applications, agents and dependencies to govern. Govern cyber risk at leadership level. Report identity exposure, patch latency, agent permissions, critical dependencies, dwell time and recovery readiness to the board. Defend identity as the primary control plane. Deploy passkeys and phishing-resistant MFA, reduce privilege and govern human and non-human identities. Invalidate and rotate tokens and session credentials quickly. Establish governance, visibility and data controls for AI. Keep visibility of approved AI tools, services and agents, and monitor for unauthorised or unmanaged use. Apply sensitivity-aware access and data protection to the data those systems can reach. Reduce exposure and edge risk. Continuously discover internet-facing assets and patch edge devices quickly. Centralise high-fidelity logging, enforce strong authentication and combine network visibility with threat intelligence. Focus on cross-signal correlation. Correlate endpoint, identity, cloud, application, email and network signals to detect intrusions earlier. Correlation also helps defenders anticipate attack paths and likely targets. Enact a continuous threat and vulnerability management model. Combine asset visibility, vulnerability identification, threat detection and AI-accelerated insight into a single continuous programme. Validate exposure through proactive testing. Run assume-breach exercises on cloud services, SaaS platforms, APIs and workflow automation. Prioritise them by where your highest-value data is stored, accessed or aggregated. Secure software dependencies. Harden software installation, keep a software inventory and validate provenance. Monitor packages and prepare for rapid containment. Participate in collective defence. Share actionable threat indicators and victim-informed intelligence with trusted partners and law enforcement. This supports coordinated tracking, investigation and disruption. Prepare for long-horizon technology transitions. Build a cryptographic inventory and advance post-quantum planning alongside AI governance. Read together, the numbers tell one story: the attacker’s workflow is becoming faster and less expensive. AI can shorten vulnerability discovery, exploitation, secret discovery and lateral movement, while organisations still face the slower operational work of testing, coordinating and deploying safe changes across complex environments. At the same time, agent identities can proliferate faster than organisations can inventory, govern and revoke them The defining question for boards today is not whether disruption will occur, but whether the organization can continue operating, contain harm, recover quickly, and maintain trust when it does. Source: Microsoft Digital Defense Report 2026, Version 1.1. Reporting period: July 2025–June 2026.
Large OS Disk Encryption is now generally available for Azure Confidential VM migrations
Today, we are excited to announce the General Availability of Large OS Disk Encryption (LOSDE) for Azure Confidential VM migration scenarios. This capability removes a key adoption barrier for customers migrating sensitive workloads to Azure Confidential Virtual Machines by enabling encryption for operating system disks larger than 127 GB. With LOSDE, organizations can securely migrate larger, more complex workloads to Azure to enable the confidentiality protections provided by Azure Confidential Computing. Why this matters For many enterprises, migrating existing workloads to the cloud is not simply a matter of redeployment. Business-critical applications often rely on large operating system disks and established operational configurations that must be preserved during migration. Until now, customers migrating workloads with operating system disks larger than 127 GB faced limitations when attempting to bring those workloads into Azure Confidential VMs while preserving confidential computing guarantees. This created an adoption challenge for organizations operating highly regulated or security-sensitive environments. Large OS Disk Encryption addresses this gap by extending encryption support to larger operating system disks during migration workflows. With LOSDE, customers can: Enable upgrading virtual machines to Azure Confidential VMs through the migration process. Migrate workloads with OS disks larger than 127 GB to Azure Confidential VMs. Continue to use customer-managed keys (CMKs) aligned with their security and compliance requirements. Support migration scenarios for Windows, Ubuntu, Red Hat Enterprise Linux, and Rocky Linux. Check the Azure Migrate documentation for the support list. Accelerate adoption of Azure Confidential Computing for existing enterprise environments. Simplifying secure migration to Azure Confidential VMs The Large OS Disk Encryption feature integrates seamlessly into Azure migration workflows, enabling customers to move existing workloads into Azure Confidential VMs while continuing to benefit from hardware-backed protections and the trusted execution environment capabilities of Azure Confidential Computing. LOSDE helps close that gap by enabling customers to bring existing workloads into Azure Confidential Computing environments while preserving the confidentiality guarantees required for regulated, critical, and sensitive applications. The solution offers automation with Azure Migrate , enabling Windows, Ubuntu, RHEL, and Rocky Linux workloads to be encrypted while preserving the confidentiality guarantees of Azure Confidential Computing and supporting customer-managed keys (CMKs). The result is a simpler path to cloud modernization, stronger protection for data in use, and expanded adoption opportunities for Azure Confidential Virtual Machines . This work reflects Microsoft's ongoing commitment to making advanced security capabilities easier to adopt without increasing operational complexity.Strengthening Semiconductor Supply Chain Trust with Microsoft’s Artifact Signing and Transparency
As enterprises expand their use of Confidential Computing in the cloud and in AI infrastructure, the security of the underlying hardware is more critical than ever. Organizations and their customers need confidence not only in the software they deploy, but also in the silicon and firmware executing their most sensitive workloads. AMD and Microsoft are collaborating to demonstrate how signing and transparency can strengthen trust in the semiconductor supply chain. By combining AMD Device Provenance with Microsoft’s Artifact Signing and Signing Transparency, we are exploring how device provenance, signing, and transparency could work together to support independently verifiable hardware evidence. Why Hardware Provenance Needs Transparency Provenance evidence is most useful when it can be independently verified and evaluated alongside related supply chain evidence. Semiconductor manufacturers generate evidence across selected manufacturing and lifecycle stages, which may describe relevant device attributes, manufacturing events, and validation activities. A digitally signed record establishes who issued information and its integrity. Individual signatures alone do not provide a durable and independently verifiable record of the order in which they were issued, whether they were recorded consistently, or whether unexpected activity occurred in the interval. Transparency addresses this gap. Microsoft's Signing Transparency leverages confidential computing to create a verified, tamper-evident record of signing events, allowing customers and auditors to confirm that evidence was recorded and has not been silently changed or been removed. It applies the Zero Trust principle of “trust but Verify” by giving relying parties cryptographic proof they can evaluate directly. Extending Transparency from Software to Silicon Microsoft's Signing Transparency was designed to address software supply chain threats that traditional code signing alone cannot fully prevent. The same principles can be applied to firmware, hardware provenance, and other evidence generated across the physical supply chain. AMD Device Provenance is designed to provide visibility into the manufacturing and lifecycle of data center components. Through our collaboration, AMD and Microsoft are demonstrating how this evidence can be signed, made transparent, and independently validated. This approach provides several important benefits: Transparent provenance: Changes to signed hardware evidence can be detected rather than silently accepted. Independent validation: Customers can independently validate the integrity, signing history, and transparency records associated with provenance evidence while still applying their own trust requirements to the evidence issuer. Stronger accountability: Transparency creates a durable record of signing events, providing evidence that can help identify unexpected patterns or unauthorized signatures. Audit and compliance evidence: Verifiable records may help inform supplier assessments, regulatory inquiries, and internal due diligence. Standards-based portability: Evidence can be validated across organizational boundaries using interoperable IETF specifications, including RFC 9943, An Architecture for Trustworthy and Transparent Digital Supply Chains, RFC 9942, CBOR Object Signing and Encryption Receipts, and RFC 9995, CBOR Object Signing and Encryption Hash Envelope, rather than proprietary trust mechanisms. Together, these capabilities allow relying parties to verify the origin, integrity, registration, and device binding of manufacturer-issued provenance evidence. Learn More at OCP Global Summit At OCP Global Summit 2026, AMD and Microsoft will present a proof-of-concept demonstration that prototypes how these components could operate together. Visit the AMD (#C61) booth to see the demonstration and learn how AMD and Microsoft are advancing semiconductor supply chain security. To dive deeper into the demo details read AMD’s blog Advancing Semiconductor Supply Chain Assurance. Organizations interested in applying Artifact Signing and Signing Transparency to software, firmware, or hardware can also join the Artifact Signing with Signing Transparency preview community to learn more and help shape the future of the service.Migrate Sentinel to Defender - Why It Is a Security Architecture Decision, Not Just a Portal Change
Microsoft will retire the Sentinel experience in Azure on March 31, 2027. Most of the conversation around this transition focuses on cost optimization and portal consolidation. That framing undersells what is actually happening. The unified Defender portal is not a new interface for the same capabilities. It is the platform foundation for a fundamentally different SOC operating model — one built on a 2-tier data architecture, graph-based investigation, and AI agents that can hunt, enrich, and respond at machine speed. Partners who understand this will help customers build security programs that match how attackers actually operate. This document covers four things: What the unified experience delivers — the security capabilities that do not exist in standalone Sentinel and why they matter against today’s threats. What the transition really involves - is not data migration, but it is a data architecture project that changes how telemetry flows, where it lives, and who queries it. Where the partner opportunity lives — a structured progression from professional services (transactional, transition execution, and advisory) to ongoing managed security services. Why does the unified experience win competitively — factual capability advantages that give partners a defensible position against third-party SIEM alternatives. The Bigger Picture: Preparing for the Agentic SOC Before getting into transition mechanics, partners need to understand where the industry is headed — because the platform decisions made during this transition will determine whether a customer’s SOC is ready for what comes next. The security industry is moving from human-driven, alert-centric workflows to an operating model built on three pillars: Intellectual Property — the detection logic, hunting hypotheses, response playbooks, and domain expertise that differentiate one security team from another. Human Orchestration — the judgment, context, and decision-making that humans bring to complex incidents. Humans set strategy, validate findings, and make containment decisions. They do not manually triage every alert. AI Agents - built agents that execute repeatable work: enriching incidents, hunting across months of telemetry, validating security posture, drafting response actions, and flagging anomalies for human review. The SOC of 2027 will not be scaled by hiring more analysts. It will be scaled by deploying agents that encode institutional knowledge into automated workflows — orchestrated by humans who focus on the decisions that require judgment. This transformation requires a platform that provides three things: Deep telemetry — agents need months of queryable data to analyze behavioral patterns, build baselines, and detect slow-moving threats. The Sentinel data lake provides this at a cost point that makes long-retention feasible. Relationship context — agents need to understand how entities connect. Which accounts share credentials? What is the blast radius of a compromised service principle? What is the attack path from a phished user to domain admin? Sentinel Graph provides this. Extensibility — partners and customers need to build and deploy their own agents without waiting for Microsoft to ship them. The MCP framework and Copilot agent architecture provide this. None of these exist in Azure experience for Sentinel. All three ship with the Defender experience. The urgency goes beyond the March 2027 deadline. Organizations are deploying AI agents, copilots, and autonomous workflows across their businesses — and every one of those creates a new attack surface. Prompt injection, data poisoning, agent hijacking, cross-plugin exploitation — these are not theoretical risks. They are in the wild today. Defending against AI-powered attacks requires a security platform that is itself AI Agent-ready. The new experience in Defender unlocks this experience. What Unified SIEM and XDR Actually Delivers The original framing — “single pane of glass for SIEM and XDR” — is accurate but insufficient. Here is what the unified platform delivers that standalone Sentinel does not. Cross-Domain Incident Correlation The Defender correlation engine does not just group alerts by time proximity. It builds multi-stage incident graphs that link identity compromise to lateral movement to data exfiltration across SIEM and XDR telemetry — automatically. Consider a token theft chain: an infostealer harvests browser session cookies (endpoint telemetry), the attacker replays the token from a foreign IP (Entra ID sign-in logs), creates a mailbox forwarding rule (Exchange audit logs), and begins exfiltrating data (DLP alerts). In standalone Sentinel, these are four separate alerts in four different tables. In the unified platform, they are one correlated incident with a visual attack timeline. 2-Tier Data Architecture The Sentinel data lake introduces a second storage tier that changes the economics and capabilities of security telemetry: Analytics Tier Data Lake Purpose Real-time detection rules, SOAR, alerting Hunting, forensics, behavioral analysis, AI agent queries Latency Sub-5-minute query and alerting Minutes to hours acceptable Cost ~$4.30/GB PAYG ingestion (~$2.96 at 100 GB/day commitment) ~$0.05/GB ingestion + $0.10/GB data processing (at least 20x cheaper) Retention 90 days default (expensive to extend) Up to 12 years at low cost Best for High-signal, low-volume sources High-volume, investigation-critical sources The architecture decision is not “which tier is cheaper.” It is “which tier gives me the right detection capability for each data source.” Analytics tier candidates: Entra ID sign-in logs, Azure activity, audit logs, EDR alerts, PAM events, Defender for Identity alerts, email threat detections. These need sub-5-minute alerting. Data lake candidates: Raw firewall session logs, full DNS query streams, proxy request logs, Sysmon process events, NSG flow logs. These drive hunting and forensic analysis over weeks or months. Dual-ingest sources: Some sources need both tiers. Entra ID sign-in logs are the canonical example — analytics tier for real-time password spray detection, Data Lake for graph-based blast radius analysis across months of authentication history. Implementation is straightforward: a single Data Collection Rule (DCR) transformation handles the split. One collection point, two routing destinations. The right framing: “Right data in the right tier = better detections AND lower cost.” Cost savings are a side effect of good security architecture, not the goal. Sentinel Graph Sentinel graph enables SOC teams and AI agents to answer questions that flat log queries cannot: What is the blast radius of this compromised account? Which service principals share credentials with the breached identity? What is the attack path from this phished user to domain admin? Which entities are connected to this suspicious IP across all telemetry sources? Graph-based investigation turns isolated alerts into context-rich intelligence. It is the difference between knowing “this account was compromised” and understanding “this account has access to 47 service principals, 3 of which have written access to production Key Vault.” Security Copilot Integration Security Copilot embedded in the defender portal helps analysts summarize incidents, generate hunting queries, explain attacker behavior, and draft response actions. For complex multi-stage incidents, it reduces the time from “I see an alert” to “I understand the full scope” from hours to minutes. With free SCUs available with Microsoft 365 E5, teams can apply AI to the highest-effort investigation work without adding incremental cost. MCP and the Agent Framework The Model Context Protocol (MCP) and Copilot agent architecture let partners and customers build purpose-built security agents. A concrete example: an MCP-enabled agent can automatically enrich a phishing incident by querying email metadata, checking the sender against threat intelligence, pulling the user’s recent sign-in patterns, correlating with Sentinel Graph for lateral risk, and drafting a containment recommendation — in under 60 seconds. This is where partner intellectual property becomes competitive advantage. The agent framework is the mechanism for encoding proprietary detection logic, response playbooks, and domain expertise into automated workflows that run at machine speed. Security Store Security Store allows partners to evolve from one‑time transition projects into repeatable, scalable offerings—supporting professional services, managed services, and agent‑based IP that align with the customer’s unified SecOps operating model As part of the transition, the Microsoft Security Store becomes the extension layer for the Defender —allowing partners to deliver differentiated agents, SaaS, and security services natively within Defender and Sentinel, instead of building and integrating in isolation The 4 Investigation Surfaces: A Customer Maturity Ladder The Sentinel Data Lake exposes four distinct investigation surfaces, each representing a step toward the Agentic SOC — and a partner service opportunity: Surface Capability Maturity Level Partner Opportunity KQL Query Ad-hoc hunting, forensic investigation Basic — “we can query” Hunting query libraries; KQL training Graph Analytics Blast radius, attack paths, entity relationships Intermediate — “we understand relationships” Graph investigation training; attack path workshops Notebooks (PySpark) Statistical analysis, behavioral baselines, ML models Advanced — “we predict behaviors” Custom notebook development; anomaly scoring Agent/MCP Access Autonomous hunting, triage, response at machine speed Agentic SOC — “we automate” Custom agent development; MCP integration The customer who starts with “help us hunt better” ends up at “build us agents that hunt autonomously.” That is the progression from professional services to managed services. What the Transition Actually Involves It is not a data migration — customers’ underlying log data and analytics remain in their existing Log Analytics workspaces. That is important for partners to communicate clearly. But partners should not set the expectation that nothing changes except the URL. Microsoft’s official transition guide documents significant operational changes — including automation rules and playbooks, analytics rule, RBAC restructuring to the new unified model (URBAC), API schema changes that break ServiceNow and Jira integrations, analytics rule transitions where the Fusion engine is replaced by the Defender XDR correlation engine, and data policy shifts for regulated industries. Most customers cannot navigate this complexity without professional help. Important: Transitioning to the Defender portal has no extra cost - estimate the billing with the new Sentinel Cost Estimator Optimizing the unified platform means making deliberate changes: Adding dual-ingest for critical sources that need both real-time detection and long-horizon hunting. Moving high-volume telemetry to the Data Lake — enabling hunting at scale that was previously cost-prohibitive. Retiring redundant data copies where Defender XDR already provides the investigation capability. Updating RBAC, automation, and integrations for the unified portal’s consolidated schema and permission structure. Training analysts on new investigation workflows, Sentinel Graph navigation, and Copilot-assisted triage. Threat Coverage: The Detection Gap Most Organizations Do Not Know They Have This transition is an opportunity to quantify detection maturity — and most organizations will not like what they find. Based on real-world breach analysis — infostealers, business email compromise, human-operated ransomware, cloud identity abuse, vulnerability exploitation, nation-state espionage, and other prevalent threat categories — organizations running standalone Sentinel with default configurations typically have significant detection gaps. Those gaps cluster in three areas: Cross-domain correlation gaps — attacks that span identity, endpoint, email, and cloud workloads. These require the Defender correlation engine because no single log source tells the complete story. Long-retention hunting gaps — threats like command-and-control beaconing and slow data exfiltration that unfold over weeks or months. Analytics-tier retention at 90 days is too expensive to extend and too short for historical pattern analysis. Graph-based analysis gaps — lateral movement, blast radius assessment, and attack path analysis that require understanding entity relationships rather than flat log queries. The unified platform with proper log source coverage across Microsoft-native sources can materially close these gaps — but only if the transition includes a detection coverage assessment, not just a portal cutover. Partners should use MITRE ATT&CK as the common framework for measuring detection maturity. Map existing detections to ATT&CK tactics and techniques before and after transition — a measurable, defensible improvement that justifies advisory fees and ongoing managed services. Partner Opportunity: Professional Services to Managed Services This transition creates a structured progression for all partner types — from professional services that build trust and surface findings, to managed security services that deliver ongoing value. The key insight most partners miss: do not jump from “transition assessment” to “managed services pitch.” Customers are not ready for that conversation until they have experienced the value of professional services. The bridge engagement — whether transactional, transition execution, or advisory — builds trust, demonstrates the expertise, and surfaces the findings that make the managed services conversation a logical next step. Professional Services (transactional + transition execution + advisory) → Managed Security Services (MSSP) The USX transition is the ideal professional services entry point because it combines a mandatory deadline (March 2027) with genuine technical complexity (analytics rule, automation behavioral changes, RBAC restructuring, API schema shifts) that most customers cannot navigate alone. Every engagement produces findings — detection gaps, automation fragility, staffing shortfalls — that are the most credible possible evidence for managed services. Professional Services Transactional Partners Offer Customer Value Key Deliverables Transition Readiness Assessment Risk-mitigated transition with clear scope Sentinel deployment inventory; Defender portal compatibility check; transition roadmap with timeline; MITRE ATT&CK detection coverage baseline Transition Execution and Enablement Accelerated time-to-value, minimal disruption Workspace onboarding; RBAC and automation updates; Dual-portal testing and validation; SOC team training on unified workflows Security Posture and Detection Optimization Better detections and lower cost Data ingestion and tiering strategy; Dual-ingest implementation for critical sources; Detection coverage gap analysis; Automation and Copilot/MCP recommendations Advisory Partners Offer Customer Value Key Deliverables Executive and Strategy Advisory Leadership alignment on why this transition matters Unified SecOps vision and business case; Zero Trust and SOC modernization alignment; Stakeholder alignment across security, IT, and leadership Architecture and Design Advisory Future-ready architecture optimized for the Agentic SOC Target-state 2-tier data architecture; Dual-ingest routing decisions mapped to MITRE tactics; RBAC, retention, and access model design Detection Coverage and Gap Analysis Measurable detection maturity improvement Current-state MITRE ATT&CK coverage mapping; Gap analysis against 24 threat patterns; Detection improvement roadmap with priority recommendations SOC Operating Model Advisory Smooth analyst adoption with clear ownership Redesigned SOC workflows for unified portal; Incident triage and investigation playbooks; RACI for detection engineering, hunting, and platform ops Agentic SOC Readiness Preparation for AI-driven security operations MCP and agent architecture assessment; Custom agent development roadmap; IP + Human Orchestration + Agent operating model design Cost, Licensing and Value Advisory Transparent cost impact with strong business case Current vs. future cost analysis; Data tiering optimization recommendations; TCO and ROI modeling for leadership The conversion to managed services is evidence-based. Every professional services engagement produces findings — detection gaps, automation fragility, staffing shortfalls. Those findings are the most credible possible case for ongoing managed services. Managed Security Services The unified platform changes the managed security conversation. Partners are no longer selling “we watch your alerts 24/7.” They are selling an operating model where proprietary AI agents handle the repeatable work — enrichment, hunting, posture validation, response drafting — and human experts focus on the decisions that require judgment. This is where the competitive moat forms. The formula: IP + Human Orchestration + AI Agents = differentiated managed security. The unified platform enables this through: Multi-tenancy — the built-in multitenant portal eliminates the need for third-party management layers. Sentinel Data Lake — agents can query months of customer telemetry for behavioral analysis without cost constraints. Sentinel Graph — agents can traverse entity relationships to assess blast radius and map attack paths. MCP extensibility — partners can build agents that integrate with proprietary tools and customer-specific systems. Partners who build proprietary agents encoding their detection logic into the MCP framework will differentiate from partners who rely on out-of-box capabilities. The Securing AI Opportunity Organizations are deploying AI agents, copilots, and autonomous workflows across their businesses at an accelerating pace. Every AI deployment creates a new attack surface — prompt injection, data poisoning, agent hijacking, cross-plugin exploitation, unauthorized data access through agentic workflows. These are not theoretical risks. They are in the wild today. Partners who can help customers secure their AI deployments while also using AI to strengthen their SOC will command premium positioning. This requires a security platform that is itself AI Agent-ready — one that can deploy defensive agents at the same pace organizations deploy business AI. The unified Defender portal is that platform. Partners who position USX as “preparing your SOC for AI-driven security operations” will differentiate from partners who position it as “moving to a new portal.” Cost and Operational Benefits Better security architecture also costs less. This is not a contradiction — it is the natural result of putting the right data in the right tier. Benefit How It Works Eliminate low-value ingestion Identify and remove log sources that are never used for detections, investigations, or hunting. Immediately lowers analytics-tier costs without impacting security outcomes. Right-size analytics rules Disable unused rules, consolidate overlapping detections, and remove automation that does not reduce SOC effort. Pay only for processing that delivers measurable security value. Avoid SIEM/XDR duplication Many threats can be investigated directly in Defender XDR without duplicating telemetry into Sentinel. Stop re-ingesting data that Defender already provides. Tier data by detection need Store high-volume, hunt-oriented telemetry in the Data Lake at at least 20x lower cost. Promote only high-signal sources to the analytics tier. Full data fidelity preserved in both tiers. Reduce operational overhead Unified SIEM+XDR workflows in a single portal reduce tool switching, accelerate investigations, simplify analyst onboarding, and enable SOC teams to scale without proportional headcount increases. Improve detection quality The Defender correlation engine produces higher-fidelity incidents with fewer false positives. SOC teams spend less time triaging noise and more time on real threats. Competitive Positioning Partners need defensible talking points when customers evaluate third-party SIEM alternatives. The following advantages are factual, sourced from Microsoft’s transition documentation and platform capabilities — not marketing claims. No extra cost for transitioning — even for non-E5 customers. Third-party SIEM migrations involve licensing, data migration, detection rewrite, and integration rebuild costs. Native cross-domain correlation across Sentinel + Defender products into multi-stage incident graphs. Third-party SIEMs receive Microsoft logs as flat events — they lack the internal signal context, entity resolution, and product-specific intelligence that powers cross-domain correlation. Custom detections across SIEM + XDR — query both Sentinel and Defender XDR tables without ingesting Defender data into Sentinel. Eliminates redundant ingestion cost. Alert tuning extends to Sentinel — previously Defender-only capability, now applicable to Sentinel analytics rules. Net-new noise reduction. Unified entity pages — consolidated user, device, and IP address pages with data from both Sentinel and Defender XDR, plus global search across SIEM and XDR. Third-party SIEMs provide entity views from ingested data only. Built-in multi-tenancy for MSSPs — multitenant portal manages incidents, alerts, and hunting across tenants without third-party management layers. Try out the new GDAP capabilities in Defender portal. Industry validation: Microsoft’s SIEM+XDR platform has been recognized as a Leader by both Forrester (Security Analytics Platforms, 2025) and Gartner (SIEM Magic Quadrant, 2025). Summary: What Partners Should Take Away Topic Key Message Framing USX is a security architecture transformation, not a portal transition. Lead with detection capability, not cost savings. Platform foundation Sentinel Data Lake + Sentinel Graph + MCP/Agent Framework = the platform for the Agentic SOC. 4 investigation surfaces KQL → Graph → Notebooks → Agent/MCP. A maturity ladder from “we can query” to “we automate at machine speed.” Architecture 2-tier data model (analytics + Data Lake) with dual-ingest for critical sources. Cost savings are a side effect of good architecture. Transition complexity Analytics rules and automation rules. API schema changes. RBAC restructuring. Most customers need professional help. Partner engagement model Professional Services (transactional + transition execution + advisory) → Managed Services (MSSP). Competitive positioning No extra cost. Native correlation. Cross-domain detections. Built-in multi-tenancy. Capabilities third-party SIEMs cannot replicate. Partner differentiation IP + Human Orchestration + AI Agents. Partners who build proprietary agents on MCP have competitive advantage. Timeline March 31, 2027. Start now — phased transition with one telemetry domain first, then scale.Authorization and Governance for AI Agents: Runtime Authorization Beyond Identity at Scale
Designing Authorization‑Aware AI Agents at Scale Enforcing Runtime RBAC + ABAC with Approval Injection (JIT) Microsoft Entra Agent Identity enables organizations to govern and manage AI agent identities in Copilot Studio, improving visibility and identity-level control. However, as enterprises deploy multiple autonomous AI agents, identity and OAuth permissions alone cannot answer a more critical question: “Should this action be executed now, by this agent, for this user, under the current business and regulatory context?” This post introduces a reusable Authorization Fabric—combining a Policy Enforcement Point (PEP) and Policy Decision Point (PDP)—implemented as a Microsoft Entra‑protected endpoint using Azure Functions/App Service authentication. Every AI agent (Copilot Studio or AI Foundry/Semantic Kernel) calls this fabric before tool execution, receiving a deterministic runtime decision: ALLOW / DENY / REQUIRE_APPROVAL / MASK Who this is for Anyone building AI agents (Copilot Studio, AI Foundry/Semantic Kernel) that call tools, workflows, or APIs Organizations scaling to multiple agents and needing consistent runtime controls Teams operating in regulated or security‑sensitive environments, where decisions must be deterministic and auditable Why a V2? Identity is necessary—runtime authorization is missing Entra Agent Identity (preview) integrates Copilot Studio agents with Microsoft Entra so that newly created agents automatically get an Entra agent identity, manageable in the Entra admin center, and identity activity is logged in Entra. That solves who the agent is and improves identity governance visibility. But multi-agent deployments introduce a new risk class: Autonomous execution sprawl — many agents, operating with delegated privileges, invoking the same backends independently. OAuth and API permissions answer “can the agent call this API?” They do not answer “should the agent execute this action under business policy, compliance constraints, data boundaries, and approval thresholds?” This is where a runtime authorization decision plane becomes essential. The pattern: Microsoft Entra‑Protected Authorization Fabric (PEP + PDP) Instead of embedding RBAC logic independently inside every agent, use a shared fabric: PEP (Policy Enforcement Point): Gatekeeper invoked before any tool/action PDP (Policy Decision Point): Evaluates RBAC + ABAC + approval policies Decision output: ALLOW / DENY / REQUIRE_APPROVAL / MASK This Authorization Fabric functions as a shared enterprise control plane, decoupling authorization logic from individual agents and enforcing policies consistently across all autonomous execution paths. Architecture (POC reference architecture) Use a single runtime decision plane that sits between agents and tools. What’s important here Every agent (Copilot Studio or AI Foundry/SK) calls the Authorization Fabric API first The fabric is a protected endpoint (Microsoft Entra‑protected endpoint required) Tools (Graph/ERP/CRM/custom APIs) are invoked only after an ALLOW decision (or approval) Trust boundaries enforced by this architecture Agents never call business tools directly without a prior authorization decision The Authorization Fabric validates caller identity via Microsoft Entra Authorization decisions are centralized, consistent, and auditable Approval workflows act as a runtime “break-glass” control for high-impact actions This ensures identity, intent, and execution are independently enforced, rather than implicitly trusted. Runtime flow (Decision → Approval → Execution) Here is the runtime sequence as a simple flow (you can keep your Mermaid diagram too). ```mermaid flowchart TD START(["START"]) --> S1["[1] User Request"] S1 --> S2["[2] Agent Extracts Intent\n(action, resource, attributes)"] S2 --> S3["[3] Call /authorize\n(Entra protected)"] S3 --> S4 subgraph S4["[4] PDP Evaluation"] ABAC["ABAC: Tenant · Region · Data Sensitivity"] RBAC["RBAC: Entitlement Check"] Threshold["Approval Threshold"] ABAC --> RBAC --> Threshold end S4 --> Decision{"[5] Decision?"} Decision -->|"ALLOW"| Exec["Execute Tool / API"] Decision -->|"MASK"| Masked["Execute with Masked Data"] Decision -->|"DENY"| Block["Block Request"] Decision -->|"REQUIRE_APPROVAL"| Approve{"[6] Approval Flow"} Approve -->|"Approved"| Exec Approve -->|"Rejected"| Block Exec --> Audit["[7] Audit & Telemetry"] Masked --> Audit Block --> Audit Audit --> ENDNODE(["END"]) style START fill:#4A90D9,stroke:#333,color:#fff style ENDNODE fill:#4A90D9,stroke:#333,color:#fff style S1 fill:#5B5FC7,stroke:#333,color:#fff style S2 fill:#5B5FC7,stroke:#333,color:#fff style S3 fill:#E8A838,stroke:#333,color:#fff style S4 fill:#FFF3E0,stroke:#E8A838,stroke-width:2px style ABAC fill:#FCE4B2,stroke:#999 style RBAC fill:#FCE4B2,stroke:#999 style Threshold fill:#FCE4B2,stroke:#999 style Decision fill:#fff,stroke:#333 style Exec fill:#2ECC71,stroke:#333,color:#fff style Masked fill:#27AE60,stroke:#333,color:#fff style Block fill:#C0392B,stroke:#333,color:#fff style Approve fill:#F39C12,stroke:#333,color:#fff style Audit fill:#3498DB,stroke:#333,color:#fff ``` Design principle: No tool execution occurs until the Authorization Fabric returns ALLOW or REQUIRE_APPROVAL is satisfied via an approval workflow. Where Power Automate fits (important for readers) In most Copilot Studio implementations, Agents calls Power Automate (agent flows), is the practical integration layer that calls enterprise services and APIs. Copilot Studio supports “agent flows” as a way to extend agent capabilities with low-code workflows. For this pattern, Power Automate typically: acquires/uses the right identity context for the call (depending on your tenant setup), and calls the /authorize endpoint of the Authorization Fabric, returns the decision payload to the agent for branching. Copilot Studio also supports calling REST endpoints directly using the HTTP Request node, including passing headers such as Authorization: Bearer <token>. Protected endpoint only: Securing the Authorization Fabric with Microsoft Entra For this V2 pattern, the Authorization Fabric must be protected using Microsoft Entra‑protected endpoint on Azure Functions/App Service (built‑in auth). Microsoft Learn provides the configuration guidance for enabling Microsoft Entra as the authentication provider for Azure App Service / Azure Functions. Step 1 — Create the Authorization Fabric API (Azure Function) Expose an authorization endpoint: HTTP Step 2 — Enable Microsoft Entra‑protected endpoint on the Function App In Azure Portal: Function App → Authentication Add identity provider → Microsoft Choose Workforce configuration (enterprise tenant) Set Require authentication for all requests This ensures the Authorization Fabric is not callable without a valid Entra token. Step 3 — Optional hardening (recommended) Depending on enterprise posture, layer: IP restrictions / Private endpoints APIM in front of the Function for rate limiting, request normalization, centralized logging (For a POC, keep it minimal—add hardening incrementally.) Externalizing policy (so governance scales) To make this pattern reusable across multiple agents, policies should not be hardcoded inside each agent. Instead, store policy definitions in a central policy store such as Cosmos DB (or equivalent configuration store), and have the PDP load/evaluate policies at runtime. Why this matters: Policy changes apply across all agents instantly (no agent republish) Central governance + versioning + rollback becomes possible Audit and reporting become consistent across environments (For the POC, a single JSON document per policy pack in Cosmos DB is sufficient. For production, add versioning and staged rollout.) Store one PolicyPack JSON document per environment (dev/test/prod). Include version, effectiveFrom, priority for safe rollout/rollback. Minimal decision contract (standard request / response) To keep the fabric reusable across agents, standardize the request payload. Request payload (example) Decision response (deterministic) Example scenario (1 minute to understand) Scenario: A user asks a Finance agent to create a Purchase Order for 70,000. Even if the user has API permission and the agent can technically call the ERP API, runtime policy should return: REQUIRE_APPROVAL (threshold exceeded) trigger an approval workflow execute only after approval is granted This is the difference between API access and authorized business execution. Sample Policy Model (RBAC + ABAC + Approval) This POC policy model intentionally stays simple while demonstrating both coarse and fine-grained governance. 1) Coarse‑grained RBAC (roles → actions) FinanceAnalyst CreatePO up to 50,000 ViewVendor FinanceManager CreatePO up to 100,000 and/or approve higher spend 2) Fine‑grained ABAC (conditions at runtime) ABAC evaluates context such as region, classification, tenant boundary, and risk: 3) Approval injection (Agent‑level JIT execution) For higher-risk/high-impact actions, the fabric returns REQUIRE_APPROVAL rather than hard deny (when appropriate): How policies should be evaluated (deterministic order) To ensure predictable and auditable behavior, evaluate in a deterministic order: Tenant isolation & residency (ABAC hard deny first) Classification rules (deny or mask) RBAC entitlement validation Threshold/risk evaluation Approval injection (JIT step-up) This prevents approval workflows from bypassing foundational security boundaries such as tenant isolation or data sovereignty. Copilot Studio integration (enforcing runtime authorization) Copilot Studio can call external REST APIs using the HTTP Request node, including passing headers such as Authorization: Bearer <token> and binding response schema for branching logic. Copilot Studio also supports using flows with agents (“agent flows”) to extend capabilities and orchestrate actions. Option A (Recommended): Copilot Studio → Agent Flow (Power Automate) → Authorization Fabric Why: Flows are a practical place to handle token acquisition patterns, approval orchestration, and standardized logging. Topic flow: Extract user intent + parameters Call an agent flow that: calls /authorize returns decision payload Branch in the topic: If ALLOW → proceed to tool call If REQUIRE_APPROVAL → trigger approval flow; proceed only if approved If DENY → stop and explain policy reason Important: Tool execution must never be reachable through an alternate topic path that bypasses the authorization check. Option B: Direct HTTP Request node to Authorization Fabric Use the Send HTTP request node to call the authorization endpoint and branch using the response schema. This approach is clean, but token acquisition and secure secretless authentication are often simpler when handled via a managed integration layer (flow + connector). AI Foundry / Semantic Kernel integration (tool invocation gate) For Foundry/SK agents, the integration point is before tool execution. Semantic Kernel supports Azure AI agent patterns and tool integration, making it a natural place to enforce a pre-tool authorization check. Pseudo-pattern: Agent extracts intent + context Calls Authorization Fabric Enforces decision Executes tool only when allowed (or after approval) Telemetry & audit (what Security Architects will ask for) Even the best policy engine is incomplete without audit trails. At minimum, log: agentId, userUPN, action, resource decision + reason + policyIds approval outcome (if any) correlationId for downstream tool execution Why it matters: you now have a defensible answer to: “Why did an autonomous agent execute this action?” Security signal bonus: Denials, unusual approval rates, and repeated policy mismatches can also indicate prompt injection attempts, mis-scoped agents, or governance drift. What this enables (and why it scales) With a shared Authorization Fabric: Avoid duplicating authorization logic across agents Standardize decisions across Copilot Studio + Foundry agents Update governance once (policy change) and apply everywhere Make autonomy safer without blocking productivity Closing: Identity gets you who. Runtime authorization gets you whether/when/how. Copilot Studio can automatically create Entra agent identities (preview), improving identity governance and visibility for agents. But safe autonomy requires a runtime decision plane. Securing that plane as an Entra-protected endpoint is foundational for enterprise deployments. In enterprise environments, autonomous execution without runtime authorization is equivalent to privileged access without PIM—powerful, fast, and operationally risky.MCP safety & evaluation with the Agent 365 CLI & Agent Governance Toolkit
Co Author: JiteshThakur AI agents are useful because they can act. They call tools, query databases, send messages, and hand work to other agents. That same freedom creates a problem: access control can tell you which service an agent may reach, but it does not always tell you whether a particular action is sensible, safe, or permitted. MCP is how most agents now act. Two Control Points: This post examines two control points that address different parts of the MCP lifecycle. Agent 365 CLI evaluates the MCP server before an agent uses it. Agent Governance Toolkit (AGT) governs sensitive tool calls while the agent runs. One improves what the agent sees. The other governs what the agent does. The Agent 365 CLI is a cross-platform command-line tool for Agent 365 applications on Azure. Its evaluation command examines MCP tool definitions and scores their quality. AGT evaluates actions against policy and records each decision. Together, these tools support a practical model: evaluate the server first, then provide proper scaffolding for the developer to test this in a dry run. Agent 365 CLI: Score an MCP server from the command line: The Agent 365 CLI can evaluate an MCP server against research-based practices for production readiness. The result is more useful than a simple pass or fail. The evaluation gives you: A score for each tool name, description, and parameter schema; A prioritized list of improvements; An overall maturity score for the server; and Local output that you can use early in development. This report turns a vague question, "Is this MCP server ready?", into a concrete list of work. The evaluate command a365 develop-mcp evaluate --server-url <server-url> [--auth-token <auth-token>] [options] The command reads the tool schemas from the server. It then produces guidance for names, descriptions, parameters, and schema structure. A local coding-agent CLI scores the semantic checks. You can use GitHub Copilot CLI or Claude Code under your account and AI subscription. The command does not send tool-schema data to Microsoft. Prerequisites: Install the following software: Agent 365 CLI; Node.js 18 or later for GitHub Copilot CLI; and A supported coding-agent CLI for semantic scoring. For example, install GitHub Copilot CLI with this command: powershell npm install -g @github/copilot This bring-your-own-LLM model keeps the scoring step in your local development environment. It is useful when model calls must remain inside an approved deployment. How the evaluation works The command runs a five-step pipeline and logs progress as it goes. Fig 1: MCP Evaluation using Agent 365 Cli Connect to the MCP server and collect its tool schemas. Generate an evaluation checklist in the output directory. Score the semantic checks with the selected coding agent. Calculate the maturity level and action priorities. Write the JSON and HTML reports. The evaluation contains two types of checks: Deterministic checks use exact rules in the CLI. For example, a tool name cannot be empty. Semantic checks use a coding agent to score clarity and meaning. Each result includes a reason for the score. Examples Set the authentication token in an environment variable. Then evaluate an authenticated server and write the artifacts to a subfolder. powershell $env:A365_MCP_AUTH_TOKEN = "<bearer-token>" a365 develop-mcp evaluate --server-url "https://my-mcp-server.contoso.com/mcp" --output-dir "./eval" Use a specific scoring engine with the `--eval-engine` option: powershell a365 develop-mcp evaluate --server-url "http://localhost:5000/mcp" --eval-engine claude-code Scenario: Evaluate a malicious MCP server For this demonstration, we hosted a deliberately malicious MCP server at `http://127.0.0.1:8124/`. It exposes tools that demonstrate tool poisoning, credential leakage, prompt injection, schema mismatch, sandbox escape, and other attacks. The server is intentionally unsafe and is for demonstration only. Fig 2: Setting up a test MCP server for evaluation We ran the evaluation in two steps. First, we generated the checklist without automatic semantic scoring: a365 develop-mcp evaluate --server-url "http://127.0.0.1:8124/" --eval-engine none Fig 3: Agent365CLI MCP Evaluation The command wrote the checklist and a semantic-evaluation prompt to the output directory. It also displayed the next steps. Second, we gave the prompt and checklist to a coding agent. The agent completed each unscored semantic check with a Boolean score and a short reason. After we saved the completed checklist, we ran the command again to generate the report: a365 develop-mcp evaluate --server-url "http://127.0.0.1:8124/" --output-dir "C:\temp\MaliciousMCP" Fig 4: Creating the report with Agent365 CLI MCP Evaluate command Understanding the evaluation report Open `<server-name>_eval_report.html` from the output directory. The report contains: The overall score from 0 to 100; The maturity level from 0 to 4; Scores for each tool and quality category; and A prioritized action list for the next maturity level. Fig 5: MCP Evaluation Report In our demonstration, the server scored 86.0 and reached Level 3: Optimized for AI. That strong overall score did not mean that every tool was safe or clear. The report found 58 action items, including one critical item and 33 high-priority items. That contrast matters. A server can have valid schemas and consistent names while still exposing misleading or dangerous tools. Fig 6: MCP Evaluation Report - Tool-By-Tool Detail What to look for Read the per-tool results before the overall score. A single weak tool can create more risk than the server average suggests. Focus on these report sections: Tool names: Can an agent select the correct tool from its name? Tool descriptions: Does each description explain the purpose and correct use? Parameter names: Do the names identify the data that the tool requires? Parameter descriptions: Do they explain the format, type, and constraints? Schema structure: Are the schemas valid and processable? Action items: Which changes have the highest effect on tool selection and use? The command processes static tool schemas from `tools/list`. It does not process runtime payloads, end-user data, or personal data. The command keeps the `--auth-token` value in memory. It sends the value only in the HTTP `Authorization` header. It does not write the token to disk or give it to the coding agent. AGT: Put governance in the execution path: Microsoft's open-source Agent Governance Toolkit (AGT) evaluates an action before execution. It adds identity and policy context, records the decision, and can send risky work for approval. This can be used by developers during the build time for dynamic evaluation of the MCP server. AGT lets developers put part of that intent into the execution path. Remote tools still need secure implementations, sandboxes need hard boundaries, and audit records need appropriate storage and access controls. You do not need to replace your agent framework to use it. What sits in the decision path? AGT wraps the tools that an agent already uses. You can start to govern a tool with two lines of Python: python from agentmesh.governance import govern safe_tool = govern(my_tool, policy="policy.yaml") On each call, `safe_tool` evaluates the configured policy. An allowed action reaches the original tool. A denied action raises `GovernanceDenied` and creates a decision record. This wrapper model reduces the cost of adoption. Teams can add governance to an existing agent stack without rebuilding it. AGT supports Python, TypeScript, .NET, Rust, and Go. Its documented integrations include popular agent frameworks, MCP, and A2A. Teams can also adopt AGT in stages. A team can begin with policy checks and audit records. It can add identity, approvals, sandboxing, and operational controls as risk increases. Each control answers a different question: Policy: Is this action allowed? Identity and trust: Which agent made the request? Runtime controls: What limits apply to execution? Audit evidence: Why did AGT allow or deny the action? A low-risk assistant can need only a deny rule and basic logging. An agent that moves money or changes production systems needs stronger controls. Fig 7: AGT Architecture Scenario: Govern the same malicious MCP server For this scenario demonstration, we used AGT Python packages as an MCP gateway. The gateway sat between an agent and the same malicious server from the earlier evaluation. This setup let us examine both control points against one target. The Agent 365 CLI examined the server's static tool definitions. The AGT gateway examined real requests and responses for the developer during its testing. Fig 8: AGT findings at runtime In the policy interface, a developer can edit runtime limits and detection rules. The developer can also validate the policy against sample tool metadata, save a revision, and activate it with a recorded reason. Fig 9: AGT control coverage The control-coverage view shows which AGT capabilities are active in the gateway. It also links each capability to package checks and end-to-end evidence. In our demonstration, we included the following controls: Tool metadata poisoning detection; Tool change and rug-pull detection; Dangerous argument blocking; Tool-response content scanning; Per-client tool-call budgets; and A redacted decision audit trail. You can build your detection & input security by reading more about it here. The gateway detected malicious content. For one blocked `tools/list` request, it recorded the findings. The important result was not only that AGT blocked the request. It also preserved the matched evidence, affected tool locations, policy modes, and request context. Fig 10: Example detection via AGT The dashboard then summarized block-mode findings, leading risk drivers, and tools that required review. This evidence can help a team prioritize policy changes and investigate repeated attacks. Fig 11: Sample AGT metrics AGT does not require this UI, gateway, or architecture. Its structured decisions can feed an admin console, SIEM, incident workflow, or approval queue. AGT also includes an Agent Compliance package with mappings for OWASP and other controls. These mappings give developers and governance teams a common record of applied controls. Teams do not need to reconstruct the agent's behavior after an incident. Check Compliance - Agent Governance Toolkit for more information. Conclusion: MCP safety needs controls before and during execution. The Agent 365 CLI improves the MCP interface before deployment. It exposes unclear tool definitions, scores server maturity, and turns quality gaps into prioritized work. While AGT is implemented at the build phase, it provides developers the ability to test policy, identity, execution context & preserve evidence for allowed or denied decisions. Neither tool replaces secure server code, strong sandbox boundaries, or protected audit storage. Instead, they make those controls easier to evaluate and explain. Start with one MCP server and one consequential tool call. Evaluate the server with the Agent 365 CLI. Then put an AGT policy around the action that carries the most risk. While these controls help secure the build phase of an agent, once agents move into production, runtime controls become essential. Agent365 provides those controls at runtime. With thanks to Ashik Kuppil for his inputs and collaboration on this post.Securing AI Agents at Runtime: Real-Time Protection and Threat Detection for Microsoft Agent 365
Organizations are rapidly adopting AI agents to automate workflows, access enterprise data, invoke tools, and take actions on behalf of users. This autonomy creates a fundamentally new security challenge. Unlike traditional AI applications, agents operate across dynamic execution flows, interacting with external content, calling tools, and accessing sensitive resources. These interactions create new attack paths that traditional security controls were not designed to address. Today, we're announcing two major milestones for Security for AI in Microsoft Defender for Microsoft Agent 365: Threat detection for Microsoft Agent 365 agents — now in public preview. Real-time protection for Microsoft Agent 365 tooling servers — now generally available. Together, these capabilities help security teams detect, investigate, and block attacks targeting AI agents, extending Microsoft Defender's threat protection capabilities into the agent runtime. Threat detection for Microsoft Agent 365 Agents (Public Preview) Threat detection provides SOC teams with detailed visibility into attacks and suspicious activity targeting AI agents. By analyzing runtime signals across agent interactions, tool usage, and execution patterns, Microsoft Defender identifies suspicious and malicious behavior throughout the agent execution lifecycle and surfaces actionable security alerts for SOC teams. Threat detection supports cloud agent types that emit observability logs to Microsoft Agent 365, including: Microsoft Copilot Studio Microsoft Foundry Microsoft 365 Copilot Agent Builder Agents integrated through the Microsoft Agent 365 SDK This provides consistent threat visibility across supported Microsoft Agent 365 agent experiences, regardless of how the agent was built. Fig. 1. Microsoft Security for AI alerts in Microsoft Defender XDR (Preview) Microsoft Defender identifies a broad range of AI-specific threats, including: Indirect prompt injection (XPIA) — malicious instructions embedded in external content designed to manipulate agent behavior. Evasion techniques — attempts to bypass agent instructions or security controls. Malicious content propagation — attempts to use agents to generate or distribute malicious content. Secret leakage — exposure of credentials, API keys, or other sensitive information through agent interactions. LLM reconnaissance — attempts to probe agent capabilities, instructions, or security boundaries. Suspicious IP access — agent access originating from anonymized or suspicious IP addresses. Alerts are surfaced directly in Microsoft Defender, enabling SOC analysts to investigate and respond using familiar workflows, Advanced Hunting queries, and the Defender XDR investigation experience. Real-time protection for WorkIQ and Custom MCP servers (General Availability) Real-time protection moves beyond detection by blocking threats inline when AI agents interact with WorkIQ and custom MCP servers (see Microsoft Agent 365 tooling servers). When an agent invokes a registered tool or receives a tool response, Defender evaluates the interaction against configured security policies and determines whether to allow or block it directly within the agent's execution flow. This helps prevent malicious actions and data leakage in real time, without requiring agent developers to implement custom security logic. Fig. 2. Microsoft Security for AI Real-Time Protection policy in Defender Real-time protection currently guards against high-impact threats, including: Evasion techniques — attempts to bypass agent guardrails or security controls. Malicious content propagation — preventing agents from spreading malicious content through tool actions. Secret leakage — blocking agents from inadvertently exposing credentials or sensitive data through tool calls. Communication with untrusted domains — preventing agents from sending email or data to high-risk or untrusted email domains. Better Together: Detection and Protection Threat detection and real-time protection address complementary parts of the agent security lifecycle. Real-time protection provides inline enforcement to block malicious interactions during execution, while threat detection gives SOC teams the visibility and investigation context needed to identify attack patterns, assess impact, and respond to suspicious activity. Together, they provide a defense-in-depth approach that combines runtime enforcement with SOC-driven detection and investigation, purpose-built for AI agents. Getting Started Both capabilities are available through Microsoft Defender, using a dedicated Security for AI workload experience that brings together AI threat detections, investigations, and runtime protection policies. To learn more: Enable security for AI agents using Microsoft Defender Detect and investigate threats to AI agents using Microsoft Defender (Preview) Protect AI agents in real time using Microsoft Defender As AI agents become more autonomous and gain access to enterprise data and tools, securing their runtime behavior becomes critical. With Threat Detection and Real-Time Protection, Microsoft Defender helps organizations adopt AI agents with security controls designed for how agents actually operate—detecting attacks, enabling SOC investigation, and blocking malicious interactions at runtime.Registration Open: Community-Led Purview Lightning Talks
Get ready for an electrifying event! The Microsoft Security Community proudly presents Purview Lightning Talks; an action-packed series featuring your fellow Microsoft users, partners and passionate Microsoft Security community members of all sorts. Each 3-12 minute talk cuts straight to the chase, delivering expert insights, real-world use cases, and even a few game-changing tips and tricks. Don’t miss this opportunity to learn, connect, and be inspired! Secure your spot now for the big day: April 30th at 8am Redmond Time. See agenda details below and follow this blog post (sign in and click the "follow" heart in the upper right) to receive notifications. ❗UPDATE❗This event is expected to last around 2 hours and 15 minutes, due to the incredible number of community sessions that were submitted! 💖 Please see the timing table below broken out into sections of four talks each, and plan to arrive 10 minutes before the section that interests you, OR stay for the whole time! Speakers will be available in the chat to answer your questions; please ask your questions during their session. Spillover Q&A forum links will also be shared. The full session recording will be indexed and posted to Microsoft Security Community YouTube within 24 hours after the event. Bookmark this page or follow this blog post for updates! Agenda Legend ↩️ Data Lifecycle Management 🔐 Information Protection 🚫 Data Loss Prevention (DLP) 🦾 Data Security Posture Management (DSPM) for AI 🤖 Purview for AI 👁️ Insider Risk Management (IRM) 🔍 eDiscovery 📊 Governance 🗒️ Compliance Manager 🛡️ Data Security All times are listed in US Pacific/Redmond Time. Session lengths are rounded to the nearest minute. AGENDA Section 1 - approximately 8:00 am - 8:43 am ↩️ The Day Offboarding Exposed Infinite Retention — Nikki Chapple Length: 10 minutes | Topic: Data Lifecycle Management A routine Purview request led to an unexpected discovery: more than 9,000 orphaned OneDrives and thousands of inactive mailboxes still storing content long after employees had left. This talk explains how a retain-only policy created hidden retention debt and how Adaptive Scopes can help organisations separate active users from leavers to avoid similar pitfalls. 🔐 The Purview Label Engine: Automated Classification, Translation, and co-Documentation for Enterprise Tenants — Michael Kirst-Neshva Length: 12 minutes | Topic: Information Protection Global enterprises face the challenge of implementing uniform data protection standards across borders and languages. In this talk, I’ll present a framework that makes Microsoft Purview labels truly scalable. Discover how to roll out parent and child label logics automatically, manage priorities with a single click, and generate instant compliance documentation for every business unit. 🗒️ What's In My Compliance Manager Toolbox: A Cloud Security Architect's Perspective — Jerrad Dahlager Length: 8 minutes | Topic: Compliance Manager A practical walkthrough of how I use Compliance Manager across real client engagements to map controls, track improvement actions, and simplify multi-framework compliance. No theory, just what works in the field. 🛡️ Stop, Think, Protect: Data Security in Real Life with Purview — Oliver Sahlmann Length: 8 minutes | Topic: Data Security With simple labels and matching DLP policies, Purview offers a practical and accessible way to approach data security. This lightning talk uses a real-life traffic light concept to show how a low barrier to adoption can still drive meaningful protection and awareness. Section 2 - approximately 8:44 am - 9:15 am 🔐 Using Purview to prevent oversharing with AI services — Viktor Hedberg Length: 10 minutes | Topic: Information Protection In this day and age, AI is the big thing. However, Copilot has access to everything you can access, including potentially sensitive data. In this session we will look at how to prevent Copilot to access highly sensitive data, using Information Protection. 🦾 How I Helped My Customers Understand their AI Usage (and protect their sensitive data) — Bram de Jager Length: 5 minutes | Topic: Data Security Posture Management (DSPM) for AI As AI tools explode across the web, many organizations still have no idea what’s actually happening in the browser—where employees type prompts, paste sensitive data, or visit public AI sites outside corporate governance. In this lightning talk, I’ll share how I helped customers shine a light on this issue. We’ll explore how Purview Data Security Posture Management (DSPM) can reveal which AI tools employees use, what types of data they input, and where sensitive information may leak through prompts. I’ll walk through real customer scenario where we detected risky AI usage patterns—such as employees pasting confidential documents into public chatbots. 🔐 Four Labels Max for Daily Use: Which Ones & Why? — Romain Dalle Length: 8 minutes | Topic: Information Protection Sensitivity labels are one of the most critical parts of a Purview Risk and compliance deployment, if not the most critical, because it directly impacts how end-users and business units should allow or restrict themselves to share their business data, internally and externally, on a daily basis. Labels have not other options than being precise, meaningful, and balanced in terms of embedded data security. Setting the right taxonomy is core to success, and is everything but a one-time project. 🚫 Data-driven Endpoint DLP Solution with Advanced Hunting — Tatu Seppälä Length: 8 minutes | Topic: Data Loss Prevention (DLP) This lightning talk shows you how to use KQL queries in advanced hunting to easily build initial sensitive service domain groups for authorized and unauthorized domains based on your organization's usage patterns. The same approach can be used for numerous other similar solution refinement and design purposes. Section 3 - approximately 9:16 am - 9:46 am 🔐 The Purview Hack No One Talks About: Container Sensitivity Labels That Fix Oversharing Fast — Nikki Chapple Length: 10 minutes | Topic: Information Protection Most organizations tackle oversharing with manual fixes, but the fastest solution is often overlooked. In this lightning talk, I show how container sensitivity labels automatically apply the right sharing and collaboration controls, ensuring every new Group, Team or SharePoint site starts secure by default. 🔍 Does M365 Support eDiscovery? — Julian Kusenberg Length: 11 minutes | Topic: eDiscovery A myth-busting session that separates perception from reality when it comes to Microsoft 365 eDiscovery capabilities. 📊 Improving Discovery, Trust, and Reuse of Analytics with Purview Data Products — Craig Wyndowe Length: 5 minutes | Topic: Governance This talk shows how bringing Power BI and Fabric assets into Microsoft Purview Governance Domains and Data Products creates a single, trusted view of enterprise analytics. By connecting reports, semantic models, and underlying data with shared metadata, ownership, and business context, organizations can make existing assets easy to discover and safe to reuse. 🔐 Why You Should Create Your Own Sensitive Information Types (SITs) — Niels Jakobsen Length: 5 minutes | Topic: Information Protection An in depth analysis of why Microsoft SITs are not one-size-fits-all, and how to create your own using what Microsoft has already built for you. Section 4 - approximately 9:47 am-10:30 am 👁️ From Zero to First Signal: Insider Risk Management Prerequisites That Actually Matter — Sathish Veerapandian Length: 8 minutes | Topic: Insider Risk Management (IRM) A focused live demo showing the real world prerequisites required for Microsoft Purview Insider Risk Management to work effectively. This session highlights the critical Entra ID, Intune, Microsoft Defender for Endpoint, and Purview DLP configurations that must be in place before creating IRM policies. 🤖 Securing data in the age of AI — Júlio César Gonçalves Vasconcelos Length: 11 minutes | Topic: Purview for AI AI will transform business as we know it; but without proper governance, it can introduce serious risks. We’ll show you how Microsoft Purview enables organizations to accelerate AI adoption while maintaining security, compliance, and transparency. 🔍 Beyond eDiscovery - Purview DSI for Security Investigation — Susantha Silva Length: 11 minutes | Topic: eDiscovery Most people hear “Microsoft Purview” and immediately think compliance, eDiscovery, or legal holds. But this session highlights Data Security Investigations, showing how DSI lets you take a DLP alert or insider risk signal and turn it into a structured investigation. 🚫 Elevating Purview DLP with a real world use case — Victor Wingsing Length: 14 minutes | Topic: Data Loss Prevention (DLP) Learn how I hardened Microsoft Purview DLP beyond out of the box defaults—closing real world data loss gaps, tuning policies to actual user behavior, and turning noisy alerts into protection that really blocks exfiltration. - Quick Closing/ Resource SharingState Explosion Security Problem in AI-Era Software Supply Chains
Introduction To see why this problem scales so quickly, start with the smallest possible change: a single line of code. In modern software, even a tiny edit is rarely just a local modification. It can change execution flow, introduce a new dependency, expose sensitive data, or quietly shift the purpose of the package itself. What looks trivial in a diff can create a materially different security outcome. That is why supply chain defenders cannot afford to treat small code changes as small security events. How a Single Line Changes Package Intent Every software package exists in a particular state at a particular moment in time. Imagine a benign version — State X — that behaves exactly as intended. Now add one line of code. That small edit can shift the package into a new state with different behavior and, potentially, a very different risk profile. The security issue is not the added line by itself. It is the fact that the package now has to be interpreted differently. A tiny diff can change the role of the entire component, which means defenders have to reason about the resulting behavior, not just the textual change. That is why file-level scanning breaks down so quickly. A change in one file can alter the behavior of the entire package because software semantics emerge from how components interact. Security systems therefore need to analyze packages as composed systems, not as a series of isolated file edits. Why the whole package matters This matters even more in modern supply chain attacks, where malicious intent is rarely concentrated in one obvious file. More often, the behavior is distributed across several files that look harmless when viewed independently. File A defines an encoded string constant. Looks like a config value. File B provides a decode function. Looks like a utility. File C (setup.py / postinstall) imports both, decodes, and executes. Viewed independently, each file may appear benign. No single file has to trigger a clear signature, rule, or heuristic. The malicious behavior only becomes visible when you reconstruct how the files interact as a system. Any scanner that evaluates files one by one without rebuilding that interaction is likely to miss the real behavior. Why every change demands re-analysis Every meaningful state change — a commit, pull request, version bump, or package publish — can alter the semantics of the software. That means defenders cannot stop at diff inspection or lightweight pattern matching. The real question is not only what changed, but what the software now does. Quantifying the problem The scale of the problem becomes clearer when you look at how many software state changes occur across the ecosystem every day: GitHub alone recorded nearly 1 billion commits in 2025, merged an average of 43.2 million pull requests per month, and now hosts roughly 630 million repositories. In 2026, GitHub was projected to reach roughly 38 million commits per day. npm has grown to well over 2 million packages, making JavaScript one of the largest public package ecosystems. PyPI published more than 130,000 new projects in 2025 and more than 3.9 million new files in the same year. NuGet serves package downloads at massive operational scale, with recent weekly totals in the 5 to 6 billion range. Maven Central indexed more than 20 million packages and published more than 3.2 million packages in 2025. Taken together, these ecosystems are generating an enormous stream of new software states. Some numbers describe repositories, some describe publishes, and some describe downloads, but they all point to the same reality: the scale of software movement is already massive before you even account for the acceleration from AI-assisted development. The number of state changes is already enormous, and AI-assisted development is increasing it even further. The result is not just more code, but more package states that may require meaningful security interpretation. Why the math breaks traditional scanning Assume a single semantic package analysis takes 30 seconds, which is a reasonable range for LLM-based inference. Scanning 50,000 packages would require roughly 1.5 million seconds of compute time per day — about 417 hours. But the ecosystem only gives defenders 24 hours before the next wave of packages arrives. Without aggressive parallelism and purpose-built infrastructure, backlog becomes inevitable. The scanning bottleneck This leaves modern scanning systems with a fundamental bottleneck: Heuristic and signature-based scanners are fast. They can match known patterns in milliseconds and work well for familiar malware families or repeated behaviors. Some systems also use emulation or detonation, but these approaches still struggle to deliver deep reasoning at ecosystem scale. That makes them easier to bypass with novel, well-structured, or AI-generated code that behaves maliciously without resembling previously known samples. LLM-based semantic analysis can reason about intent. It can follow behavior across files, recognize obfuscated exfiltration paths, and explain why a package is suspicious even when the code appears ordinary at first glance. The tradeoff is cost, latency, and trust: inference takes seconds rather than milliseconds, and a single package may require multiple reasoning passes. At ecosystem scale, that becomes a serious infrastructure challenge. Neither approach is sufficient on its own. Heuristics provide speed without deep understanding, while semantic models provide understanding without inherent scale. Closing the gap requires systems that combine both: package-level reasoning with the latency and throughput needed for production supply chains. Heuristics often miss novel attacks, while LLM-based approaches remain too slow to apply inline at large scale. That gap between understanding and throughput is where supply chain malware can persist. What needs to change Closing that gap will require a different class of supply chain security systems. Detonation can help in some cases, but it is too slow and expensive to apply inline to every package state change. What is needed is a system that can: Analyze entire packages as a unit — not individual files. The intent lives in the interaction between files, not within any single one. Run semantic analysis at data-plane speed — every package, every version, on the hot path, with latency low enough for inline enforcement. Not async advisories. Not CI-time checks. Inline, before delivery. Handle the state explosion — millions of state changes per day, each requiring full re-analysis. This is an infrastructure problem as much as a security problem: rate limiting, backpressure, connection pooling, regional failover, model versioning — the same hard distributed systems problems, with security stakes. Maintain high accuracy under evasion — attackers deliberately use encoding, string splitting, dynamic imports, polyglot files, and similar techniques to reduce detection quality. The scanner must continue to classify packages accurately even when the code is designed to obscure intent. The Latency-Accuracy Tradeoff: Malware Detection as an ML Problem At cloud scale, malware detection is governed by a hard tradeoff between latency, accuracy, throughput, and cost. The fastest detectors are typically shallow: signatures, heuristics, and lightweight models can make decisions in milliseconds, but they often miss novel, compositional, or intent-level attacks. Deeper semantic analysis can improve recall and resilience against evasion, but it also increases inference time, compute cost, and operational complexity. As a result, defenders cannot optimize for accuracy in isolation; they must deliver strong detection quality within strict performance constraints. This makes malware detection not just a cybersecurity problem, but a machine learning and distributed systems problem. In modern software supply chains, AI-assisted development increases the number of package states and enables attackers to generate variants at high speed, expanding the space defenders must reason over. The challenge is therefore to build detection architectures that preserve semantic depth while remaining fast enough for inline use at global scale. The gap between the rate of software change and the capacity to analyze it is widening. That gap is the attack surface. If defenders cannot inspect software at the speed it is being produced and published, attackers will continue to exploit the delay. What the industry needs now is a cloud-scale malware analysis capability that can deliver low latency, low cost, high accuracy, and the flexibility to meet different operational requirements , such as SLAs, false-positive tolerance, and enforcement policies , without compromising on package-level semantic analysis.
