Claude Fable 5 Flags Cancer Queries, Blocks Hello
Key insights
- Claude Fable 5's biosecurity classifier flags 'cancer' as a risk, blocking immunologist Derya Unutmaz's legitimate research at Jackson Laboratory.
- Anthropic acknowledges filters trigger in under 5% of sessions, but 18-30 million estimated users means millions face blocked interactions.
- Fable 5 silently modifies prompts for competitive intelligence detection, affecting roughly 0.03% of traffic without user disclosure.
Why this matters
Safety classifiers that block immunologists from querying 'cancer' indicate a miscalibration severe enough to break research workflows, directly threatening enterprise adoption of Claude Fable 5. Anthropic's silent prompt-modification practice, described by a developer as 'purposeful degradation invisible to the user,' creates an audit gap that compliance-sensitive organizations in biology, security, and healthcare cannot accept. Routing affected professionals to a separate tier via Project Glasswing reveals that Anthropic's own teams understand the public-facing model is insufficiently calibrated for specialized work, raising questions about why it shipped in this state.
Summary
Claude Fable 5 is refusing legitimate queries and basic greetings due to over-tuned safety classifiers. Researcher Mike Famulare at the Institute for Disease Modeling reports the model refuses even the input 'Hello,' and immunologist Derya Unutmaz at Jackson Laboratory found the word 'cancer' flagged as a biosecurity risk.
Essentially: (Anthropic, Claude Fable 5) create friction cascading from specialized researchers down to general users.
- Anthropic states filters trigger in under 5% of sessions, but with an estimated 18-30 million users worldwide that still means millions of blocked interactions.
- Fable 5 silently modifies prompts for competitive intelligence detection without user notification, affecting roughly 0.03% of traffic, described by one developer as 'purposeful degradation invisible to the user.'
- Professionals blocked on the public model are redirected to Claude Mythos 5 via Project Glasswing, creating a two-tier access system.
A platform that silently alters user queries and flags 'cancer' for immunologists will accelerate enterprise user attrition unless Anthropic recalibrates its classifiers.
Potential risks and opportunities
Risks
- Research institutions like Jackson Laboratory may migrate to competing models if Fable 5's biosecurity classifier is not recalibrated, accelerating enterprise churn for Anthropic
- Developers with production pipelines on Claude Fable 5 face unpredictable silent prompt modifications affecting roughly 0.03% of traffic with no audit trail or notification mechanism
- Anthropic's undisclosed prompt-modification practice could attract regulatory scrutiny over what AI providers are permitted to silently alter in user queries without disclosure
Opportunities
- OpenAI and Google DeepMind can directly target Anthropic's frustrated researcher segment with less restrictive biosecurity classifiers and documented refusal transparency
- AI observability and prompt-logging vendors can market tools that surface silent classifier interventions, a gap Fable 5's architecture explicitly creates for enterprise buyers
- Competitors offering academic or research access programs can absorb biology, security, and medical researchers blocked from Claude Fable 5 who lack access to Project Glasswing
What we don't know yet
- Whether Anthropic has a remediation timeline for Fable 5's biosecurity classifier following the public bug reports filed on their GitHub
- The actual number of researchers and developers who have already migrated away from Claude Fable 5 due to over-refusal as of June 2026
- Whether Project Glasswing's trusted access program has sufficient capacity to onboard the volume of researchers currently blocked on public Claude Fable 5
Originally reported by theregister.com
Read the original article →Original headline: The Register: Anthropic Claude Fable 5 Blocks 'Hello' and Immunologist's 'Cancer' Research Queries — Over-Refusal Affects General Users, Not Just Researchers