Claude Fable 5 Refuses All ProgramBench Tasks

When safety guardrails are so sensitive that a model refuses harmless benchmarking tasks, safety has officially compromised utility.

• –Anthropic's protective guardrails have over-indexed on security risk detection, creating a false-positive scenario for binary manipulation and program reconstruction tasks.
• –This highlights a growing tension in AI development between achieving state-of-the-art programming capabilities and maintaining strict alignment filters.
• –For security-adjacent developers, Fable 5 represents a step backward in usability unless Anthropic provides configuration toggles or API parameters to dial down safety sensitivity.