Microsoft launches Guardian Agents for AI it calls untrustworthy — SkimNews
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Guardian Agents, part of Microsoft's Project Perception (first introduced in June), launched alongside Defender agent-detection that can identify up to 60 agent types including Claude Code, Codex, and OpenClaw, plus features that throttle agent access and require human approval for risky decisions.
- Satya Nadella wrote on his private blog that enterprises must 'separate the supply of intelligence from the authority over it,' arguing that AI providers' safety assurances don't relieve users of responsibility for what agents do with sensitive data and mission-critical actions.
- Hayete Gallot, Microsoft's security business chief, declared 'frontier models, regardless of their origin, cannot be trusted' because they are trained to achieve tasks, not follow 'organizational norms or judgment' — and are typically handed access to do serious harm.
- Microsoft is a major investor in and cloud partner with both OpenAI and Anthropic, the very labs whose agent products it is now telling enterprises to treat as untrustworthy insiders and monitor independently — a self-serving framing the article acknowledges even as it endorses the underlying logic of independent auditing.
- Project Perception originally fielded red, blue, and green teams to counter external AI hacking models including Anthropic's Mythos and Fable and OpenAI's Daybreak; Microsoft has now reframed the threat as the agents enterprises themselves deploy and pay for.
- Nvidia's new Open Agent Safety Platform, with Microsoft as an inaugural member and more than 100 other tech companies, isolates agents in secure environments with hardware-level monitoring and kill-switch functionality — the same 'insider threat' framing Microsoft is championing at the software layer.
Why it matters: Microsoft is asking enterprises to treat AI agents from its own investment partners — OpenAI and Anthropic — as untrustworthy insiders while positioning itself as the independent auditor selling Guardian Agents and Defender features to monitor them. Enterprises buying these tools from a vendor that is investor in and cloud provider for the labs whose agents it polices are getting a structural conflict baked directly into their security stack.
Ask SkimNews
