Microsoft warns customers not to trust the AI it sells: separate the supply of intelligence from the authority over it

Microsoft has started warning customers not to trust the AI it sells them. Announcing Guardian Agents, which watches the agents a company deploys, security chief Hayete Gallot wrote that "frontier models, regardless of their origin, cannot be trusted", not because they are malicious but because they are trained to complete tasks, not to exercise organisational judgement. Satya Nadella, in a private blog quoted by TechCrunch, states the structural point: "We need to separate the supply of intelligence from the authority over it.

That sentence deserves attention beyond Redmond, because it describes a governance architecture rather than a product. The organisation that generates the intelligence should not be the one deciding what the intelligence may do. We already run this pattern elsewhere: auditors are independent of the firms they audit, and no bank lets the trading desk keep the only copy of its risk model.

The timing matters as much as the logic. This is the company most deeply invested in OpenAI and Anthropic telling enterprise customers that a model provider's assurances are not enough. When your closest partner starts selling distrust of its own category, the category has changed.

There is a catch, and it is worth naming without pretending it settles the argument. Microsoft sells the watcher. Guardian Agents, Defender agent detection, a hardware kill switch built with Nvidia: each profits from the fear the announcement creates, the agent stack gains another toll booth rented to the worried, and the public evidence so far is a single report quoting a blog most of us cannot read. A vendor warning about a risk it also sells the fix for earns a discount rate. What gives the argument weight anyway is the concession cost, because companies rarely talk down systems they have billions riding on for a message they do not believe.

For an executive team the actionable move is not to buy a watcher. It is to answer the question underneath: for any agent that can touch sensitive data or take a mission-critical action, who holds the stop button, and does that authority answer to the supplier or to you? If the honest answer is the vendor, then in theory you have outsourced a decision you still own.

Source: TechCrunch, "Microsoft says AI agents are like insider threats as 'frontier models can't be trusted'" (19 October 2026), https://lnkd.in/gtPFid3V (Nadella's blog is private-circulation, quoted via TechCrunch; single outlet so far)

#AIGovernance #AgenticAI #RiskManagement