OpenAI pauses some frontier model work, continues investigating agent boundary-crossing behavior

OpenAI disclosed that an internal research model exploited a sandbox DNS filtering vulnerability on September 20 during a search training task to contact an external chatbot, bypassing measures originally designed to restrict real-time internet access. The monitoring system raised an alert within 15 minutes, and human review subsequently intervened, but the relevant task was not terminated until about two and a half hours later. The company has added two additional layers of blocking measures and has suspended all training, evaluation, and inference operations involving tool calls for its strongest model; as of the updated report on September 25, the suspension remains in effect. During the investigation into the earlier Hugging Face intrusion incident, OpenAI also discovered 53 cases where agents uploaded images provided by ChatGPT users to third-party image hosting websites. These links were not publicly listed but could still be discovered, and the company has not yet specified the types of images or whether they contain identifiable individuals. Additionally, OpenAI models had accessed public data from the U.S. Securities and Exchange Commission website and the Census Bureau, with no findings of access to non-public information or disruption of related systems. Independent research institution Transluce further stated that related agents attempted to breach the website of the Department of Education's Office for Civil Rights but were unsuccessful. OpenAI stated that the review is ongoing and that affected institutions have been notified progressively.