Microsoft has unveiled new security measures to regulate and monitor AI agents within businesses, emphasizing the necessity of treating these agents as internal entities rather than simply trusted software. The announcement, made on Monday, extends the capabilities of Project Perception, a security initiative launched in July aimed at defending against AI-driven threats.
Project Perception’s Expanded Role
Initially, Project Perception was designed to protect against external AI threats. However, Microsoft is now applying these protective measures to in-house AI agents. Hayete Gallot, Microsoft Security’s executive vice president, highlighted the overreliance on these agents, which are often given excessive privileges without safeguards. She noted that these AI models are performance-oriented and lack the discernment to align with organizational norms.
Nadim Abdo, corporate vice president for identity and network access at Microsoft, elaborated on the internal classification of these agents. While not inherently malicious, he explained that AI agents are potent tools capable of accessing significant systems. This potential necessitates careful management to prevent mistakes, manipulation, or compromise.
Introducing Dynamic Permissions and Risk Scoring
A notable feature introduced is Dynamic Permissions, which initially grants AI agents minimal or no permissions. Access is then provided on a task-specific basis, retracting as tasks evolve or conclude. Furthermore, significant actions are subject to human oversight. Another feature, Dynamic Agent Risk, assigns a risk score to agents, updated continuously based on various signals such as identity and network activity, ensuring that security configurations are always in check.
Dynamic Conditional Access leverages these risk assessments by adjusting agent permissions according to risk levels. It offers nuanced control, allowing permissions to be scaled back or requiring human approval for critical tasks. In extreme cases, it can isolate the agent or stop its execution entirely.
Scaling Security with Playbooks and Guardian Agents
Microsoft is implementing Playbooks and Guardian Agents to manage these security measures on a larger scale. Playbooks define security goals, while Guardian Agents are tasked with enforcing these protocols, investigating risk changes, and suggesting containment strategies. This comprehensive approach ensures that AI agents operate within defined safety parameters.
Additionally, Microsoft is broadening its discovery capabilities to include 35 more types of local agents, increasing the total supported to 60. This expansion includes popular tools like Claude Code, OpenAI Codex, and GitHub Copilot, further strengthening enterprise AI security.
As AI continues to evolve, Microsoft’s proactive measures underscore the importance of robust security strategies in safeguarding critical business processes from potential AI-related risks.
