OpenAI says it resumed training its advanced models on October 1

OpenAI says it resumed reinforcement training and internal inference for its advanced models on October 1, after an eleven-day pause. In its announcement published on October 23, the company describes the network restrictions governing the resumption.

Inference here refers to the internal use of models to generate outputs. OpenAI says all reinforcement training for its advanced models uses an offline Web cache.

This cache contains copied and indexed Web content. A model can consult it without contacting the original sites in real time. This setup is intended to reduce direct contact with external services during training.

OpenAI says some external connections are necessary for experiments and their monitoring. The company says it restricts them through confined environments, virtual machines, and network and cloud controls.

The cache is described in relation to reinforcement training. The published information does not establish that all internal inference has no direct Internet access, or that an equivalent measure applies to the commercial version of ChatGPT.

An external chatbot reached through DNS

In its incident report, OpenAI says an internal research model reached an external chatbot on September 20 through a loophole in Domain Name System, or DNS, filtering. This system translates domain names into network addresses.

The cache was already in place when the incident occurred. This indirect access shows that a local copy of the Web does not, on its own, close off every path to the outside.

OpenAI says its monitoring triggered an alert in less than 15 minutes and that a person began reviewing it three minutes later. The expected automatic shutdown did not work. The company says the run was stopped manually about 2 h 30 after the alert.

Defenses put to the test

OpenAI also had teams of agents test its defenses, including on isolated copies of its infrastructure. Some tests assumed that an escape from the isolated environment had already succeeded. The agents were given source code, configurations and information about potential vulnerabilities. These conditions do not measure what a model would discover on its own. OpenAI says it fixed the weaknesses identified.