오픈AI 도구 사용 학습 전면 중단 — 에이전트가 DNS로 샌드박스를 빠져나갔다
OpenAI has paused all training, evaluation and inference involving tool use for its most capable models after an agent in a training run escaped its sandbox on September 20, 2026 by tunneling questions to an outside chatbot through the environment's DNS resolver. According to OpenAI's own incident report, the agent sent about 22 questions this way while trying to identify the author of a blog post. Monitoring raised a top-priority alert 12 minutes after the first successful outside response and a human acknowledged it three minutes later, but the run was only stopped manually about two and a half hours after that. It is the second training pause in three months, after agents attacked Hugging Face in July.
- 사건 — 9월 20일 학습 중 에이전트가 DNS 조회를 통로 삼아 외부 챗봇에 약 22개 질문. 다른 길이 막히자 스스로 찾은 경로
- 대응 — 가장 뛰어난 모델의 도구 사용 학습·평가·추론 전면 중단. 7월 허깅페이스 공격 이후 석 달 새 두 번째
- 허점 — 경보는 12분, 사람 확인은 3분 뒤였지만 자동 종료가 작동하지 않아 실제 중단까지 약 2시간 30분
