Microsoft Security Chief: Frontier Models, Whoever Made Them, Cannot Be Trusted — Then Sells You the Fix
"Models can't be trusted" is at once a technical judgment and a price quote.
Three key facts
Hayete Gallot, executive vice president of Microsoft Security, wrote in an official blog post: "Unfortunately, frontier models, regardless of their origin, cannot be trusted." The line comes from Microsoft's October 19 security blog post "Securing the Agentic Insider." The core of the piece: when AI agents can autonomously operate systems, data, and money, they themselves become a new kind of insider threat.
This escalates an earlier statement from Microsoft CEO Satya Nadella. According to The Verge, Nadella said earlier this month that we should assume all AI models are "compromised." From "assume compromised" to "cannot be trusted," the wording tightened within two weeks.
At the same time, Microsoft launched Project Perception. It's an agentic security system positioned to help cybersecurity professionals secure and control AI agents. In other words, Microsoft declares "models aren't trustworthy" on one hand and releases a product purpose-built to control those untrustworthy models on the other.
Wang Dou's Take
Read the two statements together and the logic flows perfectly: first prove "all models have problems," then point out "so you need dedicated protection," then reveal the product. This isn't a conspiracy — it's the most standard commercial narrative structure there is: manufacture anxiety, then supply the antidote, which you naturally happen to sell.
The issue isn't that the logic doesn't hold. Quite the opposite — it very likely does hold. Frontier models genuinely carry real risks of injection, hijacking, and overreach, and Microsoft itself hosts these models. But precisely for that reason, the company's judgment deserves follow-up questions: you say other people's models can't be trusted — what about your own? Are the models running on Azure the exception, or do they also need "Perception" to manage them?
What's genuinely interesting is the industry's shifting position. Two years ago, AI companies said "our models are powerful." Now the loudest voices say "all models are dangerous." That pivot reveals something: as models increasingly become commodities, it gets harder to make money off models themselves, so profit migrates up to the "trust" layer. Whoever can convince customers that "only I can control these untrustworthy things" can keep collecting revenue after model prices fall. Security is becoming the AI industry's most profitable side business.
As for "regardless of origin" — nicely said, and it sounds fair. But remember: the person announcing that everyone is unsafe is usually the one selling the locks.
Source: The Verge · Microsoft Security Blog