Anthropic의 Fable, 보안 블로그도 읽지 못할 정도로 '안전'
anthropic의 새로운 fable 모델은 멀웨어 방지를 위해 코드 리뷰 같은 무해한 보안 작업까지 차단하며, 가드레일이 지나치게 공격적이라며 연구자들이 좌절하고 있습니다.
anthropic launched fable, a limited public variant of its mythos cybersecurity model, featuring aggressive guardrails that block any input tangentially related to cybersecurity or biology. researchers, including ibm's valentina palmiotti, note the model refuses innocuous requests like code reviews or reading blogs, often defaulting to the less capable claude opus 4.8 as a fallback. the restrictions are designed to prevent the development of malware and biological weapons, addressing Anthropic's core safety concerns. however, the broad, keyword-based filtering frustrates cybersecurity professionals who cannot use the tool for defensive engineering, highlighting the difficulty in balancing open safety against utility for legitimate researchers.