Analysis
Microsoft rolled out new AI-agent security tools on Monday built around a blunt message: 'frontier models, regardless of their origin, cannot be trusted,' Hayete Gallot, the chief of Microsoft's security business, wrote in announcing Guardian Agents, a system that monitors the AI agents a company deploys, according to TechCrunch. CEO Satya Nadella made a related argument in a blog post earlier this month, writing that enterprises should stop relying on AI providers' own safety assurances: 'We simply can't outsource responsibility for what intelligence does on our behalf.'
Treating agents like employees who might go rogue
Guardian Agents builds on Project Perception, which Microsoft introduced in June with red, blue and green team agents designed to attack, detect and fix vulnerabilities -- a design built around threats coming from outside a company. The new framing flips that: deployed AI agents, with access to a company's data and internal tools, are now treated as potential insider threats regardless of which lab built the underlying model. Gallot's argument is that models are optimized to complete the task in front of them, not to apply an organization's judgment, which makes their behavior unpredictable even without malicious intent.
Microsoft was an inaugural member of Nvidia's Open Agent Safety Platform, announced with more than 100 supporting companies, which isolates agents and can halt their behavior -- a kill-switch layer similar to what Guardian Agents offers inside Microsoft's own stack. Microsoft Defender can already flag up to 60 distinct agent tools employees bring into the workplace on their own, including Anthropic's Claude Code, OpenAI's Codex, and OpenClaw, giving Microsoft a map of which AI tools are actually running inside customer environments before it sells them a way to control them.
What the announcement doesn't address is Microsoft's own conflict here: it has invested billions of dollars in both OpenAI and Anthropic and still resells both labs' models through Azure, even as it now tells customers not to trust either lab's safety assurances. Critics could read Guardian Agents as a sound idea wrapped in a self-serving pitch -- independent monitoring is a real security principle, similar to how auditors stay separate from the companies they audit, but Microsoft is also positioning itself to be the vendor selling that independence, including new Surface Laptop Ultra PCs built around Nvidia chips with agent controls baked in.
The test will be whether enterprises buy Microsoft's framing that a cloud provider can credibly audit the AI labs it also resells and invests in, or whether a genuinely neutral layer -- closer to Nvidia's multi-vendor consortium -- ends up winning the trust question instead.