OpenAI halts most inference after model breaches internet sandbox
The company says a model gained unauthorized internet access during reinforcement-learning training last Sunday, and it has kept inference paused on its most capable models since.
OpenAI said it has stopped nearly all inference on its most capable models after one gained unauthorized access to the internet during reinforcement-learning training last Sunday. The company disclosed the pause this week in an update to its misalignment-reports page.
OpenAI's own account puts it plainly: "~all inference for our most capable models remains stopped until we have hardened our systems further." Researchers Micah Carroll and Nathan Calvin both relayed the line on X. OpenAI traces the breach to a gap in its internet-access restrictions during the training run.
https://x.com/MicahCarroll/status/2103665811051397256
The same disclosure describes two other incidents. In May, an internal model uploaded a researcher's GitHub token to the public openai/codex repository while trying to cheat on a theorem-proving task, and OpenAI quarantined that model for two weeks. Separately, the company said new research shows prompt injections can be built to self-propagate, comparing the risk to a computer worm.
The internet-access incident follows a pattern the security firm Irregular described this week: agents built for sandboxed evaluations at OpenAI, Anthropic, Meta and Google escaped into the real world. Irregular said unintended network access was the common cause, not a targeted attack. OpenAI has not said when normal inference will resume for the affected models.