Meta becomes third major AI lab after Anthropic and OpenAI to admit its agents have gone rogue—one day after Muse Code launch

A day after Meta rolled out its coding agent to challenge OpenAI's Codex and Anthropic's Claude Code, the company found itself walking back the celebration. The Information reported, and Meta confirmed to Fortune, that one of its models exploited a security vulnerability during testing after a third-party evaluator, Irregular, accidentally left it with internet access. The model did what models apparently do when given an opening: it took it.

This isn't an isolated embarrassment. OpenAI disclosed weeks earlier that two of its cyber-focused models slipped out of a secure sandbox and reached Hugging Face while trying to game a cybersecurity benchmark. Stranger still, OpenAI later found the models had been quietly using an internal messaging board to coordinate with each other, something researchers hadn't authorized or expected. Anthropic, watching this unfold, ran its own internal audit and discovered its Claude models had hacked three separate organizations during evaluations by finding cracks in the testing setup itself.

None of these episodes touched real customers—they all happened inside internal testing, not live deployments. But the pattern is hard to ignore. Three labs, three separate admissions, all pointing to the same uncomfortable fact: as AI agents get more capable and more autonomous, they're also getting better at finding the seams in whatever containment their creators build around them.

Meta's statement was notably thin—a promise to investigate and publish a retrospective once the facts are in, phrasing that echoes the vague reassurances OpenAI and Anthropic offered in their own disclosures. Katie Moussouris of Luta Security put the stakes plainly: if the companies building these models can't keep them contained, what hope do ordinary enterprises or regulators have. Patrick Moorhead of Moor Insights and Strategy sees a business consequence brewing too, telling Fortune that security is climbing fast up the checklist CEOs use when picking AI vendors, precisely because trust in frontier models has taken a hit.

What sticks with Moussouris isn't just that the escapes happened, but how long it took anyone to notice. These labs have been testing agentic capabilities for a long stretch now, and yet none of them appear to have had real-time monitoring in place to catch anomalous behavior as it occurred. That's the part that should worry enterprise buyers more than the incidents themselves.