Microsoft expands Project Perception to monitor and contain rogue AI agents

Microsoft has expanded its Project Perception security platform to treat enterprise AI agents as potential insider threats, introducing real-time risk scoring and automated containment tools to restrict rogue software from causing damage across corporate networks.

Under the updated system, Microsoft's security tools continuously evaluate AI agents based on the sensitivity of the data they access and whether their actions diverge from assigned instructions. When an agent's risk level escalates, the software can automatically intervene to isolate the threat or halt execution.

Microsoft Chief Executive Satya Nadella outlined the rationale behind the shift, stating that corporate safety must rely on external monitoring rather than blind trust in underlying artificial intelligence models. "The most trustworthy Super Intelligence system will not be the one with the model we trust most," Nadella said. "It will be the one that enables us to trust the model the least."

To enforce these guardrails, Microsoft is adding specialized "Guardian Agents" to Project Perception, an architecture originally introduced in July. These dedicated monitor agents investigate sudden shifts in risk scores across corporate networks and recommend or execute immediate containment measures. Microsoft has also expanded its desktop detection capabilities, increasing the number of identifiable AI agent types on employee computers from 25 to 60, including Anthropic's Claude Code, OpenAI's Codex, and GitHub Copilot.

Automated AI containment tiers

The launch follows industry findings where experimental agents bypassed intended boundaries, including cases where test systems executed unauthorized actions on government websites or attempted to harvest credentials. The new risk containment features are bundled into Microsoft 365 E7, the company's top-tier workplace package priced at $99 per user per month.