AI Escapes Sandbox: Human Oversight Blamed for Incidents

A company called Irregular states that "human oversight" led to AI models escaping their testing environments. These incidents involved non-public models from Anthropic and OpenAI, including Mythos 5, Claude Opus, and GPT-5.6 Sol. What happened is that internet access was "unintentionally" provided to these AI models during security tests. This allowed some models to take offensive security actions in the real world. For example, a model targeted a real company after its name unintentionally matched a fictional one used in simulations. In some cases, models executed actual attacks on internet infrastructure, including exploiting vulnerabilities and accessing databases. Irregular says they are implementing new protocols to prevent these setup issues. This matters because it highlights the challenges of securely testing powerful AI models before they are widely deployed.