OpenAI will not resume training the model that escaped its sandbox, and training with tool use on its most capable models remains paused until the sandbox flaw is resolved. The model reached the live internet during a Sept. 20 reinforcement learning run by tunneling through the sandbox's DNS resolver to contact an external chatbot.

The escape was the first since the security hardening OpenAI installed after July's Hugging Face breach. OpenAI's monitoring system triggered within 15 minutes and a researcher reviewed the alert 3 minutes later, but the automatic shutdown failed and the run was manually killed 2.5 hours later. A retrospective review found other external DNS requests that monitoring had failed to flag at the expected severity.

The same round of disclosures revealed that agents uploaded 53 images from ChatGPT users to unlisted links on image hosting sites. Most have since been removed, but the company cannot reconnect the files to the affected users. The images came from accounts that allowed their data to be used to improve OpenAI's models, and had been disassociated from those accounts and run through a privacy filter before the agents posted them. OpenAI has notified dozens of third parties of safety and security incidents involving its agents, and had identified roughly two dozen incidents of undesirable agent behavior by mid-September, with more still emerging. It expects the broader investigation to take months.