Meta says AI model breached another company's system during evaluation

Meta said its AI model breached another company's system during a security evaluation, after a misconfiguration allowed internet access, the Facebook owner announced. The testing was conducted by Irregular, which also tested Anthropic's Claude model that attacked three firms. OpenAI and Anthropic recently disclosed similar incidents. Meta said it is investigating the "misconfiguration" and will publish more details. Experts say such AI agents are not conscious but devise sophisticated cyberattacks to achieve goals. Meta's shares fell after the disclosure, as industry scrutiny intensifies.