"Better to catch too much" - but defenders say it's catching everything
Anthropic pitched Fable as a public, limited window into its powerful Mythos cybersecurity model. Within days, a chorus of security researchers pushed back - not because the model is weak, but because its guardrails are so aggressive they get in the way of ordinary defensive work.
What's tripping the filters
The complaints, aired across X and Reddit, paint a picture of overly broad blocking:
- One well-known researcher said Fable rejects anything even loosely cyber-related, down to reading a blog post.
- Others reported that asking for a code review or to write secure code trips the guardrails, with the model apparently treating security-flavored phrasing as offensive work rather than software-engineering best practice.
- When triggered, Fable pauses and notes its safety measures flagged the message for cybersecurity or biology topics, then falls back to Claude Opus 4.8 - which critics say quietly downgrades the result.
The consensus diagnosis is that the system looks keyword-based, so anything in the lexical field of cybersecurity sets it off.
Why the guardrails exist
This isn't caution for its own sake. Anthropic has been vocal about the risk that frontier models accelerate malware development or software compromise, and applies similar limits to biology over bioweapon concerns. It's the same posture behind Project Glasswing, the vetted program through which it released Mythos to critical-infrastructure organizations - recently expanded to hundreds of orgs across 15 countries.
The escape hatch, and the outlook
For professionals who need fewer limits, Anthropic offers a Cyber Verification Program that approved applicants can use for security work (OpenAI runs a similar Trusted Access scheme). Even some critics are forgiving: one veteran argued that on a release this sensitive it's better to over-block and loosen later, and expected the guardrails to evolve as frontier labs work more closely with a new generation of cybersecurity companies. The episode is a neat illustration of the central tension in shipping powerful dual-use models - tune them too loose and you enable attackers, too tight and you frustrate the very defenders you're trying to empower.