Fable 5's Hidden Safety Mechanism: High-Risk Queries Silently Routed to Weaker Opus 4.8

Cybersecurity News·medium signal

Coverage of the Fable 5 jailbreak surfaced an undisclosed architectural detail: when a query trips a classifier in high-risk categories (cybersecurity, biology, chemistry, model distillation), Fable 5 reportedly reroutes the request to the weaker Claude Opus 4.8 rather than refusing outright. That graceful-degradation design is a notable pattern for anyone building safety routing into agent stacks — refusal is replaced by capability downgrade. It also became the attack surface, since the obfuscation techniques were tuned to push requests through that downgrade path.