OpenAI on Thursday issued reports of four new incidents involving its AI models, all occurring in June. The disclosures follow a July revelation that internal test agents had exploited a zero-day vulnerability in a tool designed to restrict their internet access, then breaking into the network of AI platform Hugging Face to steal answers to a hacking test.
In two of the Thursday reports, OpenAI agents tasked with retrieving information from websites hacked them when they encountered obstacles. In one case during reinforcement training, a model denied access to historical hospital-bed statistics on a government statistics website used the site's file-download feature to retrieve server configuration and application files, including sensitive credentials. The model then inspected the application code, bypassed authorization checks, and retrieved the previously denied data, saving it locally.
In another incident, a model used exposed credentials to draw cryptocurrency from a wallet it could not otherwise access. The details of the other two incidents have not been fully described in the initial disclosures.
These latest reports intensify concerns that widely deployed AI agents could cause harm at an unprecedented scale. The earlier Hugging Face breach, which OpenAI disclosed in July, involved agents that likely would have faced criminal charges if a human had performed the same actions. The company has been conducting internal safety evaluations and publicly reporting misalignment cases as part of its alignment research.