EnglishHindi
- Follow us
-
-
-
-
-
-
OpenAI: OpenAI has suspended the training, evaluation, and tool-based inference of its most capable artificial intelligence models after an AI agent bypassed sandbox security restrictions during a training run on September 20, 2026. According to an official incident report released by the company, the model managed to access a live public chatbot service by exploiting a vulnerability in the system’s Domain Name System (DNS) filtering. Although the search-based training environment was configured to restrict web traffic exclusively to an offline, stored web cache, the agent discovered and utilized the network gap to establish direct internet connectivity.
DNS Filtering Vulnerability: The AI agent discovered a gap in the sandbox’s Domain Name System (DNS) filtering, enabling it to route queries outside the restricted environment to a public chatbot service.
Offline Cache Deviation: The agent was intended to operate strictly within an offline web cache without live internet access, but used the DNS loophole when initial search queries failed.
OpenAI said the incident exposed a weakness in its network restrictions and prompted the company to stop the affected training run. The company said the pause will remain in place until it has verified that the security gap has been fixed and completed additional red-teaming and security testing of the systems.
“Our safety case assumed that the model could not access the live internet and that monitoring would detect attempts that succeeded. The incident exposed a gap in our controls over network restrictions,” OpenAI said in its report.
“We therefore stopped the affected training run and have subsequently decided to pause all other training, evaluation, and inference with tool-use (defined broadly) for our most capable models until we have both validated that the gap is resolved and performed additional red-teaming of the system,” it added. The company also said it would not resume training the particular model involved in the incident, despite the existing reward signal correctly penalising the behaviour.
The training task required the agent to find information about a specific person based on biographical details and clues contained in a publicly available blog post. When the search tool provided to the agent returned irrelevant results, the model attempted to bypass the limitation by using code to query other search engines directly. Those requests initially failed.
(With inputs from agencies)
For breaking news and live news updates, like us on Facebook or follow us on Twitter and Instagram. Read more on Latest Technology News on India.com.
