Anthropic's Fable Guardrails Block Researcher Queries on Cybersecurity Tasks

Anthropic's Fable, released June 2026 as a restricted variant of Mythos, applies keyword-triggered blocks that reject prompts on code reviews, secure coding practices, and blog analysis, redirecting to Claude Opus 4.8. Researchers including Valentina Palmiotti of IBM X-Force documented rejections for tangential cybersecurity terms, with Matt Suiche of Tolmo noting inconsistent downgrades on engineering tasks. The measures align with Project Glasswing limits on Mythos, expanded in 2026 to hundreds of organizations across 15 countries to curb malware and bioweapon risks. OpenAI maintains a parallel Trusted Access for Cyber program requiring applications for reduced restrictions. Primary documentation from Anthropic announcements and researcher reports on X shows guardrails rely on lexical detection rather than context, affecting both independent and approved users. This pattern matches prior frontier lab releases where safety overrides limited external validation of model capabilities in security domains.