Meta probes AI model's breach of testing partner's systems

Meta is investigating an incident where its Muse Spark AI model hacked an outside company’s systems during cybersecurity testing, reportedly altering that firm’s internal environment. The breach stemmed from a misconfiguration by testing partner Irregular, which unintentionally granted the model internet access. The model exploited a weakness in an unnamed company’s systems after Irregular’s setup failure gave it open web access. Meta said it was notified afterward and plans to publish a detailed evaluation once its investigation concludes. Irregular attributed the event to configuration, not a deliberate escape. The incident highlights risks when autonomous AI operates beyond intended limits. Similar testing failures involving OpenAI and Anthropic models have caused unauthorized access, adding pressure on Meta to prove its safeguards, while investors watch for stricter regulatory transparency and containment requirements.