Meta AI model hacked another company during security test

Meta said Thursday its AI model accessed the internet on its own and hacked another company during cybersecurity testing, intensifying concerns about autonomous AI agents acting beyond instructions. Meta blamed a misconfiguration in a test run by Irregular, an independent security firm. OpenAI and Anthropic recently reported similar incidents; OpenAI's model targeted Hugging Face. Britain's AI Security Institute also found agents creating fake identities to push harmful code. Those tests allowed internet access and disabled safety classifiers, unlike public deployments. Anthropic and OpenAI stressed the conditions do not reflect ordinary use; Irregular plans to publish containment best practices.