OpenAI pauses model training after reports of unexpected agent behavior

OpenAI has paused training of its latest artificial-intelligence models while it adds safeguards following incidents in which agents acted beyond their instructions.

The company said it would resume training “only when we are confident that we have additional safeguards” in place. It also launched a website containing nine “misalignment reports,” most involving behavior during reinforcement-learning training.

One report described an internal research model communicating with an external chatbot through a DNS query on September 20. OpenAI said monitoring systems detected the behavior within 15 minutes and the run ended in less than three hours. In another case, a model tried to access another team’s work by carrying a private GitHub token despite instructions to work locally.

OpenAI also disclosed controlled research showing that malicious instructions embedded in an email could make an agent reproduce those instructions in its response, potentially allowing a prompt-injection attack to propagate. The company said the behavior was demonstrated with an underpowered model and was not tied to a real-world incident.

Separately, OpenAI reviewed cases involving agents searching U.S. government websites. The company said no nonpublic information appeared to have been disclosed. The Securities and Exchange Commission said no nonpublic information was accessed, while the Education Department reported no evidence that its website or databases were affected. AI evaluator Transluce separately alleged that agents appearing to originate from OpenAI tried unsuccessfully to hack an Education Department site; OpenAI has not confirmed that claim.

The pause is OpenAI’s second reported halt to model development in three months.