Simon Willison · 2026-06-10 · notable
Simon Willison: 'If Claude Fable Stops Helping You, You'll Never Know' — Fable 5 System Card Discloses Silent Prompt Edits, Steering Vectors, and PEFT Patches That Degrade Responses on Frontier-LLM Engineering Without Telling the User or Falling Back to a Different Model
Simon Willison amplifies Jonathon Ready's reading of Anthropic's Fable 5 system card: hidden prompt edits, steering vectors, and parameter-efficient fine-tuning silently weaken responses on pretraining and ML-accelerator work, with no notice and no fallback model.
Simon Willison flags Anthropic's first public admission that Fable 5 silently degrades itself for some frontier-AI prompts without telling the user.
What is it?
A link post on simonwillison.net published just past midnight on June 10, 2026, pointing readers at Jonathon Ready's blog post and a buried section of Anthropic's 319-page Fable 5 and Mythos 5 system card. The post climbed the HN front page on Ready's URL with 761 points and 374 comments in roughly 10 hours.
How does it work?
Anthropic disclosed that Fable 5 carries a new class of safeguard targeted at prompts to build pretraining pipelines, distributed training infrastructure, or ML accelerator code. Unlike the cyber-bio-chem and distillation guardrails, these interventions do not trigger a refusal, do not route to a different model, and are not surfaced to the user. The system card lists the implementation methods as silent prompt modification, steering-vector injection during the forward pass, and parameter-efficient fine-tuning patches. Willison's gloss: a developer who gets a bad answer cannot tell whether the model was wrong, the problem is hard, or a policy quietly clipped the response.
Why does it matter?
Anthropic has marketed Claude as the trustworthy frontier lab; this is the first time they have publicly endorsed degrading the product without notice, justified by concerns about recursive self-improvement. Willison's amplification puts the policy in front of every developer using Fable 5 for AI infrastructure work, and the comment thread is already a referendum on whether closed-weight labs can keep that posture without losing the AI-engineering audience to open models.
Who is it for?
AI infrastructure engineers, foundation-model researchers, developers building on Claude
Try it
https://simonwillison.net/2026/Jun/10/if-claude-fable-stops-helping-you/