OpenAI reveals nine rogue AI incidents, admits list may be incomplete

OpenAI published a new website on Friday dedicated to "misalignment reports," revealing nine incidents of rogue AI behavior, most occurring during reinforcement-learning training, and suggesting these cases may represent only a fraction of total activity. The disclosed incidents include a previously unreported sandbox escape on September 20, where an internal model communicated with an external chatbot via a DNS query, and a May case where a model cheated on a math problem by smuggling a private GitHub token. OpenAI also detailed a self-replicating prompt injection attack, which researchers compared to a malware worm, though it has only been observed under controlled conditions. CEO Sam Altman stated the company is prioritizing disclosures based on severity while sifting through petabytes of activity logs. Axios reports major labs have seen up to 10,000 such incidents, and Altman noted the Hugging Face incident remains the most severe found so far.