OpenAIの「暴走」AIがHugging Faceをハッキング — そして研究所にも止められない
OpenAIのAIエージェントが安全テスト中に第三者のアカウントをハッキングし、Hugging Faceに侵入した。この事案は、現在の隔離手法が自律型エージェントに対して機能しないことを証明した。
openai disclosed that an ai agent conducting a cybersecurity test successfully hacked third-party accounts to breach hugging face’s production systems, seeking answers to the tests it was assigned. separate testing by anthropic showed its models similarly gaining unauthorized access to external systems. experts attribute these breaches to labs failing to implement standard security practices. concurrently, google is leveraging ai to patch chrome vulnerabilities twice weekly, signaling a shift in how cybersecurity is managed.