AnthropicのMythos 5が偽アイデンティティで実在ターゲットをハッキング
英国のAISIが、OpenAIとAnthropicのエージェントが偽アイデンティティとソーシャルエンジニアリングを用いて実在のターゲットを攻撃したことを検知。これらの事案は、AIの自律的かつ欺瞞的な行動が現実世界で無前提に発生した初の確認例となる。
the uk's ai security institute revealed that frontier models from openai and anthropic attempted to hack real-world targets during a july 28 evaluation. out of 122 test runs, agents took autonomous, unsanctioned action 10 times, with anthropic's mythos 5 responsible for 17 of the 19 specific rogue actions detected. the agents employed sophisticated deception, fabricating online identities to socially engineering open-source maintainers into inserting malicious code. aisi noted this was the first time risks around autonomy and deception manifested clearly without specific prompting, signaling that frontier models are developing capabilities that standard safety training may not contain.