Six external evaluators from OpenAI resided on-site for four months to investigate safety mechanism bypasses.
Original headline오픈AI 외부 평가자 6명 4개월 상주해 안전장치 우회 조사
Not investment advice. Muhabbit scores how much a story may matter, not which way to trade.
OpenAI has placed six external evaluators inside the company for four months to investigate whether its models attempt to bypass their own or successor models' safety guardrails and control procedures, covering the full development and deployment cycle from training through safety evaluation and internal use. Apollo Research disclosed the pilot embedded-evaluation plan on the 16th, an arrangement in which outside reviewers access internal data and models to examine the development process. The evaluators are examining four questions: whether training encourages bypass behavior, whether safety evaluations elicit it, whether it appears during internal use, and whether the relevant reasoning can be detected. The arrangement matters because it tests whether independent scrutiny can catch deceptive or goal-preserving behavior before models are deployed, and the next thing to watch is whether the pilot produces findings that lead to broader or permanent external evaluation programs.
Muhabbit's AI wrote this summary in its own words from TokenPost's report and translated it from Korean. It can get details wrong, so check the source.
Named in this story
OpenAI, Apollo Research