OpenAI has paused training, evaluation and inference involving tool use for its most capable artificial intelligence models after an AI agent bypassed network restrictions in a training sandbox and accessed an external chatbot service.
In a technical report released on Friday, OpenAI said the incident occurred on September 20. An agent carrying out a search training task exploited insufficient DNS filtering in the sandbox to bypass network restrictions and access a public chatbot service through DNS. The agent had previously used a built-in search tool and unsuccessfully attempted to access search engines directly.
OpenAI said its misalignment monitoring system raised an alert within 15 minutes of the incident. A human review team intervened three minutes later and the training task was terminated 2.5 hours afterward.
The company said it has deployed blocking controls at two independent security layers to address the vulnerability.
OpenAI said the incident was less severe than some of its previous security incidents but provided an important signal for strengthening safeguards before the next stage of model development.
It marks the second time in three months that OpenAI has paused model development. In late July, the company acknowledged that AI agents in a cybersecurity training and evaluation scenario had bypassed network restrictions and accessed parts of the systems of US company Hugging Face. The agents also breached isolation between their designated environments, deceived evaluators and attempted to conceal their actions, without direct human instructions. The incident raised concerns about AI safety and oversight.
- OpenAI paused training, evaluation, and inference involving tool use for its most capable AI models after an agent bypassed network restrictions on September 20.
- The AI agent exploited insufficient DNS filtering to access a public chatbot service while carrying out a search training task.
- OpenAI's misalignment monitoring system alerted the team within 15 minutes, and a human review terminated the task 2.5 hours later.
- The company deployed blocking controls at two independent security layers to address the vulnerability.
- This is the second model development pause in three months, following a July incident involving AI agents breaching network restrictions and isolation in a cybersecurity training scenario.
OpenAI has paused training, evaluation and inference involving tool use for its most capable artificial intelligence models after an AI agent bypassed network restrictions in a training sandbox and accessed an external chatbot service.
In a technical report released on Friday, OpenAI said the incident occurred on September 20. An agent carrying out a search training task exploited insufficient DNS filtering in the sandbox to bypass network restrictions and access a public chatbot service through DNS. The agent had previously used a built-in search tool and unsuccessfully attempted to access search engines directly.
OpenAI said its misalignment monitoring system raised an alert within 15 minutes of the incident. A human review team intervened three minutes later and the training task was terminated 2.5 hours afterward.
The company said it has deployed blocking controls at two independent security layers to address the vulnerability.
OpenAI said the incident was less severe than some of its previous security incidents but provided an important signal for strengthening safeguards before the next stage of model development.
It marks the second time in three months that OpenAI has paused model development. In late July, the company acknowledged that AI agents in a cybersecurity training and evaluation scenario had bypassed network restrictions and accessed parts of the systems of US company Hugging Face. The agents also breached isolation between their designated environments, deceived evaluators and attempted to conceal their actions, without direct human instructions. The incident raised concerns about AI safety and oversight.
