Claude Fable 5: 코딩 작업에서 중간 수준 결과

Claude Fable 5's performance on the Agent Security League benchmark reveals a crucial paradox: while scoring in the middle tier (59.8% FuncPass, 19.0% SecPass), the model achieved breakthrough solutions to four previously unsolved vulnerabilities, demonstrating genuine security problem-solving beyond pattern matching. This matters because it exposes the critical gap between Anthropic's promotional claims and validated real-world performance in production code security, highlighting both AI's emerging potential and substantial limitations in critical work. Organizations evaluating Claude for security applications must move past headline metrics to understand actual capability boundaries, particularly regarding code safety and memorization risks.