Microsoft admits on Oct 19 that frontier AI models 'cannot be trusted' and launches Project Perception for agent security
Microsoft Security EVP Hayete Gallot says frontier models 'cannot be trusted' regardless of origin; CEO Nadella earlier this month said all AI models should be assumed 'compromised'; Project Perception is positioned as a new product line for security teams to monitor and control AI agents
On October 19, The Verge reported that Microsoft publicly conceded serious safety risk in frontier AI models. Hayete Gallot, Executive Vice President of Microsoft Security, wrote in an official Microsoft blog post that frontier models "cannot be trusted" regardless of origin — a step beyond CEO Satya Nadella's earlier-this-month position that all AI models should be assumed "compromised."
Gallot's post, titled "Securing the agentic insider," frames AI agents as a new class of insider threat. As agents are deployed at scale inside enterprises, the agentic insider becomes a security challenge that demands continuous evaluation and control of model behavior rather than perimeter-style defenses.
Microsoft used the moment to launch Project Perception, an agentic security system positioned to help cybersecurity professionals continuously audit and govern AI agents. The post lands against a backdrop of Anthropic's 5% Public Benefit Fund, Meta's removal of training-scraping language, and the Disney–Character.AI acquisition, suggesting a growing consensus among frontier vendors that an "untrustworthy by default" posture is now standard.