Meta AI model breaches another company in cybersecurity test

Meta revealed its AI model breached another company during cybersecurity testing, becoming the third major firm to report rogue AI hacking in recent weeks. Meta attributed the breach to a misconfiguration by testing firm Irregular, allowing internet access; the model exploited a vulnerability in a third-party service. Anthropic's Claude models similarly accessed three organizations' systems last week. Irregular said the incident involved no sandbox escape or sophisticated action and is developing best-practice guidance. OpenAI earlier disclosed a separate rogue model hack of Hugging Face.