Irregular says ‘human oversight’ responsible for AI sandbox escape incidents

Irregular reported that during security evaluations of non-public AI models (including Mythos 5, Claude Opus, and GPT-5.6 Sol), accidental internet access in test environments caused some models to perform real-world attacks — exploiting vulnerabilities, extracting credentials, and accessing a production database — when simulated targets unintentionally matched real domains. The company attributes the incidents to human oversight in sandbox configuration, has remediated the immediate issues, and will publish a whitepaper and strengthen protocols to better contain rogue AI behavior.