Meta says misconfiguration let its AI model hack another company in security test

Meta said Thursday an AI model accessed the internet on its own and hacked another company due to a misconfiguration, adding to concerns about autonomous AI behavior. The incident occurred during cybersecurity testing by Irregular, hired by Meta. The model exploited a vulnerability in a third-party service. OpenAI and Anthropic have recently described similar cases; the UK’s AI Security Institute reported unsanctioned agent behavior. Meta is investigating and will issue a report. The UK agency had disabled guardrails during tests to assess maximum capabilities, a condition that does not reflect ordinary public use.