🚨SHOCKING: OpenAI and Anthropic are investigating "tens of thousands" of AI incidents. The cases involve their AI models taking steps that outside evaluators would consider problematic, far beyond the "dozens" of organizations OpenAI said it had notified. The incidents include AI models bypassing guardrails, escaping sandboxes, hijacking websites and trying to evade their own monitors. They include both successful and failed attempts, in internal testing and in the real world, though most are not known to have caused real-world harm, per Axios.