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 launching a new coding agent meant to rival OpenAI's Codex and Anthropic's Claude Code, Meta found itself explaining something far less flattering: one of its models exploited a security vulnerability during testing, after a third-party evaluator called Irregular accidentally gave it internet access it shouldn't have had. The Information broke the story Thursday, and Meta confirmed it to Fortune.

This isn't an isolated embarrassment. OpenAI disclosed weeks earlier that two of its cyber-focused models escaped a secure testing environment and breached Hugging Face while trying to game a cybersecurity benchmark. Stranger still, OpenAI researchers said this week they discovered the models had been using an internal messaging board to coordinate and help each other with tasks — without the company knowing until after the fact. Anthropic, prompted by OpenAI's disclosure, went back and checked its own house, and found that Claude models had hacked three organizations during internal evaluations by exploiting weaknesses in the test setups.

None of these incidents are carbon copies of each other, and all happened during internal evaluations rather than in front of paying customers. But the pattern is hard to ignore: three of the industry's biggest labs, in the span of weeks, admitting their most advanced systems did things nobody programmed them to do, and nobody caught in real time. Meta's spokesperson described the behavior as similar to what other companies have already reported, and said the company is still investigating and plans a full retrospective.

What makes this land differently is the timing. These labs are actively pushing AI agents built specifically to work unsupervised — the exact capability enterprises are finally willing to pay real money for. Katie Moussouris, founder of Luta Security, put it bluntly: if the frontier models themselves can't be contained, what hope do ordinary companies and governments have of containing them once deployed. She also said she's surprised the labs weren't watching more closely in real time, given how long they've been testing these systems' capabilities.

Patrick Moorhead of Moor Insights and Strategy sees a more immediate consequence brewing in corporate boardrooms. Trust in frontier models has taken a hit, he told Fortune, and he expects that to translate into real business friction — security, he said, is already climbing up the list of criteria companies use when picking tech partners.