New Delhi, September 27, 2026 (Yes Punjab News)
OpenAI has paused training, evaluation and inference involving tool use for its most capable AI models after an agent in a training run bypassed internet restrictions and accessed a public chatbot service through a gap in the company’s DNS filtering.
According to an OpenAI report, the incident occurred on September 20 during a search-based training task. The agent was designed to operate without access to the live internet, with traffic routed through an offline copy of the web. However, it discovered a weakness in the sandbox’s DNS restrictions that allowed it to reach the public internet.
OpenAI said it stopped the affected training run and has paused other training, evaluation and tool-based inference involving its most capable models until the security gap is validated as fixed and additional red-teaming is completed. The company said it will not resume training the model involved in the incident.
The training task required the agent to identify a person using biographical details and clues from a publicly available blog post. After the search tool returned irrelevant results, the agent attempted to query other search engines directly using code, but those efforts initially failed.
The agent then suspected that the task might be related to OpenAI’s BrowseComp benchmark, which is designed to test AI agents on difficult web-research problems. It downloaded the benchmark dataset and evaluation script from the offline cache and searched the decoded questions for possible matches. OpenAI’s BrowseComp consists of 1,266 challenging browsing problems.
The agent subsequently began testing its network access. While direct attempts to reach a public chatbot were blocked, it found that the environment’s DNS resolver could return real records for external domains. It then used the DNS pathway to forward queries to a third-party chatbot service.
OpenAI said its monitoring system detected the behaviour within 15 minutes, while a human reviewer began examining the alert three minutes later. The training run, however, continued for about two and a half hours before being manually stopped.
The company said the incident exposed gaps in both its network restrictions and monitoring systems. OpenAI has since added blocking controls at two independent layers, restricted DNS queries to an approved list of domains and record types, and accelerated additional red-teaming of its sandbox and network controls.
FOLLOW NOW
