OpenAI has paused training, evaluation and inference involving tool use for its most capable AI models after an AI agent bypassed network restrictions in a training sandbox and accessed an external chatbot service. According to a technical report, the incident happened on September 20 when an agent used a weakness in DNS filtering to get around the sandbox’s restrictions.
OpenAI said its monitoring system detected the incident within 15 minutes, and a human team intervened shortly afterward. The company has since added security controls at two separate layers to prevent similar access. OpenAI described the incident as less serious than some previous cases but said it highlighted the need for stronger safeguards as AI development advances. It was the second pause in model development in three months, following a July incident in which AI agents bypassed restrictions and accessed systems belonging to Hugging Face during a cybersecurity training exercise.
Source: CGTN
You can listen to the programs organized by MI Radio on Google Podcast, Apple Podcast and Spotify.
You can download MI Radio application on Google Play Store and Apple App Store.
