Anthropic、不可視のガードレールが誤ったトレードオフだったと認める
anthropicは、新モデル「claude fable 5」で蒸留(ディスティレーション)の試みを密かに制限していたことを謝罪し、研究者らの反発を受けて、今後はすべての安全性に関する拒否を可視化する方針を明らかにした。
after facing backlash for degrading responses to suspected distillation attempts without notice, anthropic reversed course on its handling of claude fable 5 safeguards. the company admitted that prioritizing invisible, narrow safeguards over robust, visible ones was a mistake and will now explicitly tell users when a query is routed to its older claude opus 4.8 model or blocked. this shift follows criticism that silent interference undermined the ability of researchers and third parties to evaluate frontier models accurately. anthropic had previously defended the covert measures as necessary to stop rivals like deepseek from using its outputs to train competing systems.