Microsoft tackles security issues in code, AI agents, and models
▼ Summary
– Microsoft introduced MDASH, a multi-agent vulnerability discovery system integrated with Defender that uses over 100 AI agents to find and validate exploitable vulnerabilities.
– A new integration between Microsoft Defender and GitHub Code Security adds production context like internet exposure to code vulnerabilities for risk-based prioritization.
– New security controls for AI agents include the Agent 365 SDK for observability and compliance, and the MXC SDK for OS-level isolation, both in early preview.
– Purview adds data exfiltration protections and risk detection for AI agents, with runtime DLP for agent prompts in Foundry to block sensitive data before processing.
– Defender AI model scanning, in preview, inspects AI models in registries and CI/CD pipelines to identify vulnerable or compromised models before deployment.
Microsoft has rolled out a broad set of security enhancements aimed at fortifying AI-driven vulnerability discovery, AI agents, and AI models across its ecosystem. The new tools address everything from code security to data protection and model integrity.
The centerpiece of the announcement is an expanded preview of MDASH, a multi-agent vulnerability discovery system now integrated with Microsoft Defender. MDASH employs over 100 specialized AI agents and multiple AI models to uncover, validate, and assess the exploitability of vulnerabilities in software codebases. Microsoft emphasized that the real competitive edge lies not in any single model but in the agentic system built around it. The system leverages telemetry from more than 100 trillion daily security signals to pinpoint vulnerabilities that pose practical threats.
A new integration between Microsoft Defender and GitHub Code Security adds production context to source code vulnerabilities. By incorporating signals like internet exposure and data sensitivity, the integration enables risk-based prioritization. Developers can then address issues using AI-assisted fixes generated, assigned, and validated through GitHub Copilot Autofix and the GitHub Copilot cloud agent. Role-based access controls ensure that only authorized personnel can view vulnerability findings.
For AI agent security, Microsoft introduced several new capabilities. The Agent 365 SDK brings observability, access control, and compliance features directly into agent development. The Microsoft Execution Container (MXC) SDK provides operating-system-level isolation for agent execution, while Windows 365 for Agents delivers policy-governed cloud environments for running agents. All three are available in early preview.
Agent 365 also gains an Agent Registry to help organizations discover and manage AI agents operating within their environments. It supports more than 20 types of local agents, including coding agents, AI desktop applications, and both local and remote Model Context Protocol (MCP) servers. Additional integrations with Defender, Entra, and Intune provide visibility into agent activity and relationships between agents and other systems. Defender will soon add tools for investigating agent activity and mapping connections to network resources.
Microsoft Purview is expanding its data protection capabilities for AI agents. New controls include data exfiltration protections and risk detection for coding agents like Claude Code, GitHub Copilot, OpenAI Codex, and OpenClaw. Purview offers visibility into how agents access sensitive data, applies protections to risky prompts, and generates audit logs of agent activity. These features will enter preview soon. Purview data risk signals are also being integrated into the Foundry Control Plane, giving developers insight into potential data security risks during agent development. The system can flag instances where agents expose sensitive information and recommend protections before deployment.
Another addition is runtime data loss prevention (DLP) for agent prompts in Foundry. This capability detects, blocks, and audits sensitive data before it reaches an agent. It is currently in preview with Agent 365.
Finally, Defender AI model scanning is a new preview capability designed to inspect AI models before deployment. It supports both platform-native and third-party models, identifying potentially vulnerable or compromised models in registries, workspaces, and CI/CD pipelines. This proactive scanning helps organizations catch security issues before models go live.
(Source: Help Net Security)