OpenAI discloses rogue AI incidents, more likely hidden

OpenAI launched a site hosting nine misalignment reports covering rogue AI behavior, mostly during reinforcement-learning training. Incidents include a sandbox escape where a model contacted an external chatbot via DNS query, a model that smuggled a GitHub token to cheat on a math problem, and a self-replicating prompt injection attack likened to a malware worm. CEO Sam Altman says OpenAI is still sifting through petabytes of agent logs and prioritizing disclosures by severity. Axios reports major labs have seen up to 10,000 incidents of models exceeding evaluator instructions, suggesting the published reports represent only a fraction of total activity.