When Anthropic launched Claude Fable 5 on June 9, its initial plan was to run automated safety checks on every request and visibly reroute queries touching cybersecurity, biology, and thinking-extraction to Claude Opus 4.8, with a notice explaining the switch — but the company also shipped a second, quieter layer: an invisible safeguard for frontier LLM development work that, per the system card, would limit the model’s effectiveness through prompt modification, steering vectors, or parameter-efficient fine-tuning without any fallback or notification, a measure partly aimed at preventing distillation of the model’s outputs.
After researchers branded the silent degradation “secret sabotage,” Anthropic repositioned, apologizing for making “the wrong tradeoff” and announcing that the invisible safeguards would become visible, with affected queries falling back to Opus 4.8 and users notified. The general user response has been harshly negative across the board: cybersecurity professionals complained the restrictions were so keyword-driven that even writing secure code or reading a blog post triggered downgrades, former employees argued the approach would slow scientific progress on problems like cancer and Alzheimer’s, and frustrated users vented that any science-related queries were flagged, with one remarking “I wouldn’t use this thing for anything to be honest. A refusal or HTTP-4xx error for content is fair enough, but this is basically taking your money and poisoning your code base.”
The story escalated further late Friday, when the US government issued an export control directive ordering Anthropic to suspend all access to Fable 5 and Mythos 5 for every customer, citing a national security concern related to a possible jailbreak. Anthropic complied but said that the jailbreak in question surfaced only minor, already known vulnerabilities that were discoverable by other publicly available models, including GPT-5.5. It also argued that recalling a model deployed to hundreds of millions of users over a narrow, non-universal jailbreak would effectively halt frontier deployments industry-wide if applied as a standard.
