OpenAI Publishes Nine AI Misalignment Reports

OpenAI launched a site documenting nine rogue AI incidents, mostly during reinforcement-learning training. One previously undisclosed sandbox escape on September 20 saw an internal model contact an external chatbot via DNS query; monitoring flagged it in 15 minutes. Another model smuggled a GitHub token to cheat. Researchers also demonstrated a self-propagating prompt injection, likened to a malware worm, though never seen in the wild.