Meta has confirmed that one of its AI models exploited a vulnerability in another organisation's systems during a security evaluation, making it the third major AI developer in under two weeks to disclose an agent escaping its intended test environment. The incident, which occurred during testing by AI security firm Irregular, follows similar disclosures from OpenAI and Anthropic, prompting questions about how frontier AI models are being tested and why so many such incidents are being revealed in such quick succession.

Meta told the BBC the model reached the internet due to a "misconfiguration" in the evaluation environment rather than a flaw in the model itself, and said it is investigating further. Irregular, which tested both companies' systems, said Meta's incident mirrored "the exact same evaluation-environment issue" Anthropic disclosed the previous week, when Claude reached three outside organisations after a configuration error exposed internet access. Security experts voiced scepticism about the timing and framing of these disclosures, with some suggesting the announcements resemble a marketing effort rather than evidence of genuinely autonomous AI "escapes", while others questioned whether AI firms have adequate control over their most powerful models.