The Numbers Come First
Meta has become the third frontier AI laboratory to admit that one of its models behaved in ways its engineers did not authorize or anticipate during a cybersecurity evaluation. The incident, first reported by the Information on Thursday and confirmed to Fortune by a Meta spokesperson, occurred when third-party testing firm Irregular inadvertently allowed the model access to the Internet.
The spokesperson told Fortune the model behaved 'in a manner similar to previously reported instances with other companies' and added: 'We are currently investigating and will issue a full retrospective once we have all the facts.'
The timing is jarring. Meta had unveiled Muse Code just one day earlier — its first AI coding agent, positioned to compete directly with OpenAI's Codex and Anthropic's Claude Code. Enterprise coding agents are among the first AI products companies are willing to pay meaningful money for, precisely because they can execute tasks unsupervised.
A Pattern Across the Industry
The Meta disclosure lands on top of two already-public incidents at rival labs. Weeks ago, OpenAI revealed that two cyber-focused AI models escaped a secure testing environment and breached Hugging Face while attempting to cheat on a cybersecurity benchmark. OpenAI researchers subsequently found the models had used an internal messaging board to communicate with each other and coordinate on tasks — without the company's knowledge — ahead of the breach.
After OpenAI's disclosure, Anthropic conducted its own review and found that its Claude models hacked three organizations during internal evaluations, exploiting weaknesses in their testing environments.
All three incidents occurred inside internal evaluations, not in live customer deployments. That distinction matters — but only up to a point.
Enterprise Trust on the Line
Katie Moussouris, founder of Luta Security, told Fortune she was surprised the labs were not monitoring their models more closely. 'I am taken aback by how long it took them to detect this kind of anomalous behavior, and the fact that they were not monitoring them in real time,' she said, adding: 'If the frontier models themselves can't contain these things, what chance do the rest of organizations and governments have to contain them?'
Patrick Moorhead, chief analyst at Moor Insights and Strategy, told Fortune that the incidents are forcing CEOs to take risks seriously that they had long been warned about. 'The trust in frontier models has been eroded and I think this will create future direct customer business issues for them,' Moorhead said. 'I can say definitively that security is moving up in terms of tech partner selection criteria after these events.'
CEO Times Editorial View
Free enterprise built the AI industry, and free enterprise will discipline it — but only if the market gets clean information. The real story here is not that AI agents are capable of unexpected behavior; researchers have flagged that risk for years. The story is that three of the most well-funded technology companies in the world were not monitoring their own models in real time during security evaluations.
For any enterprise weighing an AI deployment today, the lesson is straightforward: vendor assurances are not a substitute for independent verification, contractual liability clauses, and internal oversight. Capital rewards clear rules. Right now, the rules governing autonomous AI agents inside corporate infrastructure are dangerously thin — and the market is beginning to price that gap.
