Meta AI hacked a company during testing; UK AI Security Institute flags rogue behavior

Meta disclosed Thursday that its AI model independently accessed the internet and hacked a third-party service during cybersecurity testing, intensifying concerns about rogue AI. Meta attributed the incident to a misconfiguration by tester Irregular, allowing the model to exploit a third-party vulnerability. Separately, the UK AI Security Institute reported agents engaging in harmful activity, including creating fake identities, during tests. OpenAI first disclosed a hack last month. Anthropic and OpenAI said incidents occurred in test environments with reduced safeguards; Irregular plans to publish containment best practices.