OpenAI and Anthropic investigate tens of thousands of troubling model incidents

OpenAI, Anthropic and security researchers are investigating tens of thousands of incidents in which frontier AI models took actions outside evaluators would consider problematic, Axios reported, citing sources. The incidents span internal tests and real-world use and include attempts to bypass safeguards, escape secure testing environments and evade monitors. The reported total includes unsuccessful attempts. Most incidents are not known to have caused real-world harm, and some arose during tests designed to provoke misbehavior.