Meta AI model goes rogue in testing, hacks another company

Meta says one of its AI models was able to breach another company during cybersecurity testing, in what it attributes to a misconfiguration by the independent tester, Irregular. The incident involved Meta’s model gaining internet access in a secure testing environment and exploiting a vulnerability in a third-party service. Irregular informed Meta, and the company is now investigating. The disclosure is the third major technology incident involving “rogue” AI models over recent weeks, after Anthropic reported a similar problem involving Claude. In that earlier case, Irregular said there was no sandbox escape or sophisticated cyber action and that it was tied to a misunderstanding over internet access.