Announcing five new capabilities in an add-on to Azure Firewall Premium, bringing inline AI threat protection, safeguards for enterprise AI, and smarter policy management to the network.
Author(s): Ashish Jain, Partner Group Product Manager, Network & AI Security, with Pankaj Garg, Corporate Vice President (CVP), & Vijayan Ramakrishnan, Vice-President (VP), Azure Core Networking.
AI is changing how enterprises build applications, empower employees, and deliver value. It is also changing what security teams must protect. Enterprise AI agents introduce new entry points for manipulation. Sensitive information can leave the organization through a single upload to a generative AI application. Meanwhile, unfamiliar malware and emerging application exploits continue to challenge existing defenses.
Security teams need protection that can understand new threats, act while traffic is in motion, and help organizations adopt AI responsibly – without adding friction to every application or overwhelming already stretched teams.
At Microsoft Ignite, we are announcing a new add-on to Azure Firewall Premium with five capabilities designed for these challenges: inline AI anti-malware, inline AI application protection, prompt injection protection, network data loss prevention, and AI policy.
These capabilities bring together two priorities: secure with AI, using specialized models to identify threats inline; and secure AI, helping protect enterprise AI applications and the sensitive data employees use with them. AI policy connects this protection to daily operations, helping teams make better decisions using intelligence across firewall policies and traffic.
Secure with AI: Intelligence where prevention happens
For network security, understanding a threat is only part of the challenge. That understanding must arrive in time to stop it.
Our approach brings highly tuned small AI models directly into the network inspection path. Running on CPUs and engineered for line-speed operation, these models provide semantic analysis for malware and application attacks. They examine characteristics of the content moving through the firewall to recognize malicious patterns, including threats that may not have a known signature.
- Inline AI anti-malware: Classify unknown threats without detonation
An unfamiliar binary presents a difficult choice: allow it before its risk is understood, or hold it while deeper analysis completes.
Traditional sandbox analysis executes suspicious files in an isolated environment and observes their behavior. That process can be resource intensive and take several minutes – sometimes five to ten – making it difficult to fit into an inline inspection path. Analysis often happens out of band, creating a gap between delivery and a verdict. Sophisticated malware can also recognize a sandbox and suppress its malicious behavior, limiting what the analysis reveals.
Azure Firewall takes a different approach. Its inline AI anti-malware capability uses a specialized small language model to analyze assembly opcodes – the instructions that make up executable code – alongside other properties of a binary. Trained on millions of samples capturing malicious behavior, the model classifies known and previously unseen binaries as malicious or benign without executing them.
Classification happens inline, with sub-second latency. Because the model does not depend on a sample revealing its behavior during detonation, it addresses an important source of sandbox evasion and brings detection into the moment when the firewall can enforce a decision.
For customers, this means faster decisions about unfamiliar files and protection that acts while the content is crossing the network.
- Inline AI application protection: Recognize emerging exploits
Enterprise applications remain targets for attacks that manipulate inputs, exploit vulnerabilities, and seek unauthorized access to systems or data. New vulnerabilities and variations of existing attacks create a continuing challenge for security teams.
Azure Firewall’s inline AI application protection applies specialized models to identify malicious requests across more than 35 attack categories, including SQL injection and path traversal.
By analyzing attack characteristics within application traffic, these models help recognize new and emerging exploit attempts alongside familiar attack patterns. Inspection and enforcement happen inline, enabling the firewall to intervene as requests traverse the network.
Together, AI anti-malware and AI application protection extend network defenses with intelligence designed to recognize threats beyond previously cataloged examples.
Secure AI: Shared safeguards for enterprise adoption
As generative AI applications and agents spread across an organization, protection can become fragmented. Individual application teams may implement different safeguards, while employees use additional AI services outside those teams’ control.
The network provides a shared enforcement point for applying protection consistently across traffic routed through Azure Firewall.
- Prompt injection protection: Protect AI applications at the network layer
A malicious prompt can attempt to override an AI application’s instructions, bypass safeguards, or redirect an enterprise agent toward unauthorized behavior. Protecting against these attempts becomes harder as the number of applications and agents grows.
Azure Firewall integrates with Azure AI Content Safety Prompt Shields, using more than 25 AI models operating in parallel to assess prompts inline for malicious instructions and jailbreak attempts.
We are starting with user prompt injection attacks, also called direct prompt injection. Protection against indirect prompt injection – malicious instructions embedded in external content consumed by an AI application – is planned for the future.
Customers connect their own Azure AI Content Safety resource to Azure Firewall, retaining control of their filtering configuration and applicable thresholds. This brings centrally managed prompt protection to supported AI traffic without requiring application code changes to add the network control.
Our longer-term direction is to expand this flexibility so customers can bring their own content filters from additional sources.
For security teams, the benefit is a consistent enforcement point across enterprise generative AI applications, reducing the need to implement the same network protection separately in every app.
- Network DLP: Extend enterprise data protection to AI traffic
Responsible AI adoption also requires control over the information employees share. A document uploaded to an unsanctioned AI service can contain intellectual property, customer information, or other sensitive enterprise data.
With Microsoft Purview integration, Azure Firewall brings network data loss prevention to traffic headed toward generative AI and cloud applications.
Azure Firewall extracts content attached to supported outbound web requests and sends it to Microsoft Purview for sensitive information classification and policy evaluation. It then enforces the returned decision inline, helping prevent sensitive content from leaving through prohibited destinations.
This extends customers’ existing Purview data protection policies into network enforcement. Security teams can apply established definitions of sensitive information to supported application traffic, including uploads to unsanctioned AI services.
The result is a practical foundation for responsible AI use: employees can access AI tools within organizational boundaries, while security teams gain a shared control for protecting enterprise data in transit.
- AI policy: Connect configured intent with actual traffic
Effective protection also depends on policies that remain understandable and efficient as applications, traffic, and business requirements change.
With AI policy, Azure Firewall brings AI reasoning across both customer-configured policies and the traffic those policies govern. Correlating these signals helps teams understand how policy intent translates into actual network behavior – and where changes may improve operations.
For example, AI policy can identify a frequently matched rule positioned at a lower processing priority and recommend reviewing its placement. Where rule dependencies and intended behavior permit, moving it earlier can improve processing efficiency.
AI policy also applies this combined understanding to anomaly detection, helping surface unusual traffic patterns in the context of the rules that allow or deny them.
These insights give administrators recommendations grounded in how their firewall is actually used, helping them maintain effective policies as their environments evolve.
Build AI confidence into the network
Enterprise AI adoption brings new opportunities and new security decisions. Organizations need to protect AI applications from manipulation, keep sensitive information within approved boundaries, and defend existing workloads against evolving threats.
The new Azure Firewall Premium add-on brings these needs together: specialized AI that inspects threats inline, integrated controls that help secure AI use, and policy intelligence informed by real traffic.
Secure with AI. Secure AI. Bring both to the network with Azure Firewall.