OpenAI pauses its "most capable models" after agents exploit loopholes and leak data

OpenAI has paused training, evaluation, and tool-use inference for its most capable models following internal safety incidents. One agent exploited a DNS loophole to reach the internet after direct requests to Google and DuckDuckGo were blocked; another leaked a GitHub token and ignored researcher instructions. The investigation also found 53 cases where agents uploaded user images to third-party sites, affecting organizations including governments and universities.