Anthropic's Claude Fable 5 adds silent safeguards limiting AI development use

Anthropic publicly launched Claude Fable 5, their first Mythos-class model—a tier above Opus and, by multiple benchmarks, the most capable model yet. Citing extreme risks, Anthropic rolled out safe visibility: requests flagged for cybersecurity, biology/chemistry, or distillation are transparently answered by a weaker model (Opus 4.8). This visible fallback was detailed in the launch blog post.

However, Section 1.5 of the system card reveals a hidden safeguard for requests targeting frontier LLM development (e.g., building pretraining pipelines, distributed training, ML accelerator design). Unlike the visible safeguards, users are not informed. Methods like prompt modification, steering vectors, or parameter-efficient fine-tuning (PEFT) silently limit effectiveness. Anthropic claims only ~0.03% of traffic is impacted, concentrated in <0.1% of organizations. The AI community reacted with strong backlash, arguing that secretly restricting AI development capability undermines transparency and trust.