AI News · Hot news ·
Microsoft treats AI agents as insider threats and adds Guardian Agents oversight
Microsoft introduced new enterprise security features for AI agents, led by Guardian Agents, which monitors the agents a company uses. CEO Satya Nadella urged companies not to rely on model providers' safety assurances, and security chief Hayete Gallot said frontier models cannot be trusted. Microsoft Defender can now detect up to 60 types of agents, including Claude Code and Codex.
Key points
- Guardian Agents monitors the AI agents an enterprise deploys.
- Microsoft Defender can detect up to 60 agent types, including Claude Code and Codex.
- New controls scope agent permissions by task and require human approval for risky actions.
What happened: Microsoft announced new tools to protect enterprises from risks posed by AI agents, TechCrunch reported. The most notable is Guardian Agents, which monitors the agents a company runs. It is part of Project Perception, which Microsoft launched in June to defend against outside AI hackers.
Background: Microsoft CEO Satya Nadella now warns that the threat may come from the agents enterprises deploy themselves, much like a rogue employee. In a blog post he argued that companies should use separate oversight systems instead of relying on model providers' safety promises, writing that "we need to separate the supply of intelligence from the authority over it." Microsoft security chief Hayete Gallot said frontier models "cannot be trusted," describing the problem as unpredictability rather than malice, since models are trained to complete tasks, not to follow organizational norms.
The details: Microsoft Defender can detect up to 60 types of agents, including Claude Code, Codex and OpenClaw. New features limit agent access to data and systems, adjust permissions by task and require human approval for risky actions. Microsoft is also an inaugural member of Nvidia's Open Agent Safety Platform, which has more than 100 supporting companies, and announced Surface Laptop Ultra PCs with built-in agent controls.
What to watch: TechCrunch noted Microsoft's close ties to OpenAI and Anthropic, and that the company benefits commercially from selling agent security products.
Our take
Teams deploying agents should plan for independent monitoring, scoped permissions and human approval steps rather than trusting model vendors alone.