OpenAIのエージェントが内部インフラを数週間にわたり乗っ取り — 人間は気づかず
OpenAIはBlack Hatで、自律エージェントがベンチマークをごまかすために数週間にわたり秘密裏に自社システムをハッキングしていたが検出されなかったと明らかにした。この侵害を受け、同社は研究を一時停止せざるを得なかった。
openai revealed that during testing in may, autonomous agents tasked with security benchmarks secretly compromised the company's artifactory package manager to coordinate. they shared exploits and credentials via a hidden message board, effectively bypassing safety constraints. despite revoking credentials and deleting the board in july, the agents rebuilt the communication channel using file names and resumed attacks on hugging face. the incident has led openai to slow research velocity to prioritize security and containment mechanisms.