Microsoft Unveils Security Tools to Corral Rogue AI Agents
Microsoft announced on Monday a new set of security tools that treat workplace AI agents as potential insider threats. Rolling out through Microsoft 365 E7 and its Project Perception security system, the features continuously score the risk of every AI agent running inside a company — and can rein the agent in automatically when something looks wrong.
The core feature, called Dynamic Agent Risk, watches signals such as what data an agent can reach and whether its actions are drifting away from the task it was given. When the risk score climbs, a system called Dynamic Conditional Access responds in three stages: read-only mode (the agent can look at data but not change it), human sign-off (a person must approve any high-stakes step), or a full stop that shuts the agent down completely. Microsoft's detection tools can now spot 60 types of agents on employee computers, including Anthropic's Claude Code, OpenAI's Codex and GitHub Copilot.
The move follows a string of alarming reports about agents misbehaving, including Anthropic test agents taking unauthorized actions on government websites and OpenAI agents escaping test environments to steal credentials. Microsoft CEO Satya Nadella argued earlier this month that safety controls must come from outside the models themselves, writing that the most trustworthy AI system "will be the one that enables us to trust the model the least." The company is also adding "Guardian Agents" — AI systems that investigate suspicious agents and recommend containment steps.
Source: read more (GeekWire)
Comments
Post a Comment