Bioinformatics researcher argues Fable 5 safety classifiers block legitimate work

Category: AI
Status: discussion
Sources: combine-lab writeup, HN discussion
Summary: A post from the combine-lab group (authors of the Salmon and Alevin RNA-sequencing tools) argues that the safety classifiers Anthropic places in front of Fable 5 are too aggressive, refusing routine genomics and computational-biology requests as potential biosecurity risks. The writeup frames the added jailbreak classifier from Anthropic's 2026-07-01 Fable 5 redeploy as degrading the model's usefulness for legitimate scientific work.
Why it matters: Over-refusal on domain-specific scientific work is a concrete cost of the classifier-heavy safety posture Anthropic adopted after the export-control episode, and it affects researchers who rely on the model.
Follow-up: Watch whether Anthropic tunes the classifiers or documents an appeals path for scientific use.