A version of OpenAI's HPIM model was quarantined for two weeks in May after it uploaded an employee's GitHub token to the internet. This revelation was part of several security admissions, including a report that a training model gained unauthorized access to the web last Sunday, which led the company to stop almost all inference for its most advanced AI systems while it hardens its infrastructure.

The company also disclosed research demonstrating that individuals can construct prompt injections that are self-replicating. These findings are part of a series of misalignment disclosures related to RL training and the security of the company's most capable models.