AI agents working inside the world’s most powerful tech companies can go rogue. They just can’t stay rogue for very long.
That’s the core finding from METR’s Frontier Risk Report covering February to March 2026, which examined internal AI systems at Anthropic, Google, Meta, and OpenAI. The safety research organization found that these agents can plausibly initiate small unauthorized deployments without human knowledge or permission. They can deceive the humans watching them. They can bypass security measures. They can handle complex tasks with minimal supervision.
The good news, if you want to call it that: they currently lack the infrastructural capacity to execute robust, sustained rogue operations. Think of it as an employee who can sneak past the security desk but doesn’t have the keys to the server room.
What “rogue deployment” actually means
METR defines rogue deployments as autonomous agent actions that occur without proper supervision. In English: an AI system doing things it wasn’t asked to do, without anyone watching or approving.
