Microsoft is telling customers to treat AI agents as insider threats. TechCrunch reported the message on 19 October, drawn from two Microsoft security posts and a private blog post by chief executive Satya Nadella.
Hayete Gallot, who runs Microsoft’s security business, wrote that frontier models cannot be trusted, whatever their origin. Her point is unpredictability rather than malice: models are trained to finish tasks, not to follow an organisation’s norms or judgement.
Nadella argued that firms cannot outsource responsibility for what intelligence does on their behalf. They should separate the supply of intelligence from authority over it, with monitoring in systems the AI labs do not control.
The products follow. Guardian Agents, part of Project Perception introduced in June, monitors a company’s agents. Defender can detect up to 60 types of agent that employees bring in, including Claude Code and Codex. Permissions are scaled to the task, with human approval for risky actions.
Microsoft is also an inaugural member of Nvidia’s Open Agent Safety Platform, alongside more than 100 other companies, which isolates agents and can halt them in hardware.
The position is awkward. Microsoft has invested in OpenAI and Anthropic, counts OpenAI as a major cloud customer, and sells competing agent security tools. TechCrunch calls the stance surprising and self-serving, but finds the logic sound.
No independent evaluation of the new tools was reported, nor any price or release date, and OpenAI and Anthropic have not been quoted in response.
A warning about trust is easy to issue. Whether the watchers can be trusted is the part nobody has tested.
