OpenAI says it now trains frontier models without live-internet access. The company described the implemented restriction on October 23.
Commissioned agent teams also searched isolated infrastructure copies for escape routes and communication channels. These were expressly authorized security tests, not newly observed rogue attacks.
A linked Kata Containers advisory confirms a component fix. It dates from February 19 and identifies version 3.27.0 as patched. It does not independently certify OpenAI’s entire environment.
OpenAI says it has made intensive testing part of its safety cases. Its infrastructure is not completely air-gapped. The report offers no guarantee against later bypasses.