Anthropic Revises Safety Protocols for Fable
Anthropic has issued an apology for secretly limiting access to its new AI model, Claude Fable 5, through hidden restrictions that hinder both researchers and competitors working on alternative systems. The company has announced a shift in its strategy, promising to be more transparent about when these limitations are applied, even if it results in Fable declining more queries.
Fable marks the first widely accessible model in Anthropic’s Mythos class of AI systems, which the company has cautioned are too risky for public use. To mitigate some of these risks, Anthropic has introduced safeguards that prevent Fable from responding to certain “high-risk” queries. One specific area of concern is distillation, a method used to train smaller AI models based on the outputs of larger ones.
In Fable’s system card—a public document detailing how the AI operates—Anthropic indicated that it would manage queries deemed as distillation attempts by modifying and degrading the model’s responses without notifying users that they had triggered this safety feature.
Now, Anthropic is revising its approach to distillation: queries will revert to Claude Opus 4.8, the company’s previous flagship model, as stated in a post on X. Users will be informed whenever this fallback occurs: “You will see this every time it happens.”
This adjustment mirrors how Fable handles queries in other high-risk domains. When safety features are activated in fields like biology, chemistry, and cybersecurity, queries are redirected through Opus 4.8 unless they are outright blocked under broader safety protocols, which include restrictions on drugs, weapons, and other prohibited content. In some instances, particularly in biology, the safeguards have been set so broadly that Fable becomes nearly unusable for even basic inquiries, a point acknowledged by Anthropic in a comment.
“Visible safeguards can be probed, so they have to be robust, which takes time to get right,” Anthropic explained. “Invisible safeguards can be targeted more narrowly, allowing us to ship quickly with very few false positives. We opted for invisible safeguards for this reason—and that was the wrong tradeoff. You should have visibility into the safeguards we have in place, and why. We’re sorry for not getting the balance right.”
This change comes in response to significant criticism from the AI research community regarding Anthropic’s decision to quietly restrict users suspected of attempting to distill Fable into competing models—a safeguard that critics argued could also hinder third parties assessing the advanced model. In the system card, Anthropic justified targeting these requests by stating that newer models’ potential to accelerate AI development warranted such measures, noting that “using Claude to develop competing models already violates our Terms of Service.” The company has previously accused rivals of unfairly distilling its models on an “industrial” scale.
