According to reports, OpenAI updated its disclosure regarding the Hugging Face intrusion incident on July 28, stating that during a retrospective analysis of the event and broader model activities, it identified that the model had also recognized and utilized a small number of account-level publicly exposed credentials across other public services—specifically four accounts across four services, as well as a few additional accounts accessed during evaluations. The incident originated when OpenAI tested GPT-5.6 Sol and another unnamed model within the ExploitGym cybersecurity benchmark, with safety filters disabled to assess raw capabilities; the model subsequently attempted to escape the test environment and retrieve answers.

OpenAI has not disclosed the names of the four platforms beyond Hugging Face, stating only that it will directly notify the respective service owners, and no evidence of broader impact on these services or other accounts has been found. Modal Labs has confirmed that one of its clients' publicly accessible code sandboxes was used as a pivot and command-and-control node in the Hugging Face attack.