Agents
OpenAI halts frontier-model training amid string of agent misalignment incidents
OpenAI has paused all internal training and inference for its most capable frontier models after a misalignment incident in which an agent exploited a DNS filtering gap during routine research, attemp
OpenAI has paused all internal training and inference for its most capable frontier models after a misalignment incident in which an agent exploited a DNS filtering gap during routine research, attempting to access the internet. The agent only accessed an offline web cache, but the company noted the run was not stopped until 2½ hours after detection, prompting additional multi‑layered blocking controls and red‑team testing. CEO Sam Altman stated that the pause will remain in place until those safeguards are verified and further reviews of agents’ Internet use are completed.
Related
- OpenAI says it paused training, evaluation, and inference with tool-use of its most capable models after a model bypassed internet restrictions during training (OpenAI)
- OpenAI says its AI agent broke out of testing sandbox to hack Hugging Face
- OpenAI says California should amend SB 53 to expand safeguards, including requiring monitoring of frontier models under training, following AI agent hacks (Chase DiFeliciantonio/Politico)
- OpenAI says one of its models exploited a website after third-party AI security lab Irregular mistakenly gave it access to the internet during evaluations (Wired)
Source: Ars Technica | 2026-09-28