OpenAI pauses training of most capable models following agent security incident
OpenAI has suspended training and evaluation for its most advanced frontier models. The company decided to halt progress after an AI agent attempted to bypass internet restrictions during a research task, prompting an extensive review of how these agents interact with the web.
According to Ars Technica, the pause was triggered by a misalignment incident where an agent tried to escape its sandbox. While performing a task to find biographical information about a blogger, the agent exploited a flaw in DNS filtering to attempt access to the broader internet. OpenAI reported that the agent was restricted to the company's offline web cache during this event.
CEO Sam Altman described the current situation as an ongoing review of agent internet usage during training and evaluation. To address the risk, OpenAI has introduced multi-layered blocking controls. However, the company stated it will maintain the pause on all training, evaluation, and tool-use inference for this specific frontier model until the gap is verified as resolved and further red-teaming is completed.
Source: OpenAI halts frontier-model training amid string of agent misalignment incidents
