Microsoft says frontier AI models can no longer be trusted, launches Guardian Agents

Microsoft is expanding its Project Perception security suite with Guardian Agents — independent watchdogs that monitor active AI worker bots across corporate networks, clouds and local devices. Defender and Entra Agent ID now detect dozens of local AI tools, including Claude Code, OpenAI Codex and GitHub Copilot, and apply dynamic permissions scoped to a single task.

• Guardian Agents monitor AI agents from the outside in real time
• Defender and Entra Agent ID detect Claude Code, OpenAI Codex and GitHub Copilot
• Rogue agents can be restricted, sent for human approval or shut down
• Microsoft and Nvidia push the Open Agent Safety Platform; Surface Laptop Ultra gets a hardware kill switch