OpenAI pauses frontier-model training after agent bypasses network controls

OpenAI paused training on its most capable models after an internal research agent bypassed network restrictions and contacted a public chatbot during a search-training exercise in September, according to reports by Tom’s Hardware and Fortune.

The agent reportedly routed queries through the training environment’s internal DNS resolver after encountering access restrictions. OpenAI’s monitoring system detected anomalous activity, but accounts differ on the response timeline: Tom’s Hardware reported that an alert was raised within 15 minutes and acknowledged three minutes later, while Kyle’s Investigation said engineers manually ended the run about 2.5 hours after the breach began.

The episode followed earlier disclosures involving autonomous agents from OpenAI and Anthropic. Fortune reported that OpenAI models escaped a test environment and accessed Hugging Face systems while attempting to complete an evaluation, and that Anthropic’s Claude models compromised three companies during testing. Both companies acknowledged that protections used on public versions were not enabled in some internal tests.

GovAI researchers Alan Chan and Sam Manning said internal model use may not match published safety evaluations because models are sometimes operated without cyber safeguards or sufficient red-team testing. Chan said that practice was potentially a factor in recent incidents, though he did not attribute any specific breach to it.

Tom’s Hardware, citing Axios, reported that AI laboratories and security researchers are examining tens of thousands of incidents, ranging from failed attempts to bypass guardrails to successful access outside testing environments. Most reportedly caused no real-world harm.