Microsoft Launches Guardian Agents, Warns Frontier AI Models Cannot Be Trusted

Microsoft introduced a new security feature called Guardian Agents on Monday alongside updates to Microsoft Defender that can detect up to 60 types of AI agents brought into the workplace by employees, including Claude Code, Codex, and OpenClaw. Microsoft security business chief Hayete Gallot warned that frontier models are unpredictable and trained to complete tasks rather than follow organizational judgment, while CEO Satya Nadella compared deployed enterprise AI agents to human insider threats. Guardian Agents expands Project Perception, which Microsoft launched in June with red, blue, and green agent teams in response to security and hacking models like Anthropic's Mythos and Fable and OpenAI's Daybreak. The new enterprise controls restrict agent access to sensitive systems, dynamically adjust permissions based on specific tasks, and require human approval when an agent attempts risky decisions. Microsoft also joined more than 100 tech companies supporting Nvidia's Open Agent Safety Platform to isolate and stop agents via hardware, following the launch of Nvidia-powered Surface Laptop Ultra PCs with built-in agent security controls. "Unfortunately, frontier models, regardless of their origin, cannot be trusted." — Hayete Gallot, Microsoft Security Business Chief.