原文 · 未翻译
AI
News
Anthropic
Anthropic apologizes for invisible Claude Fable guardrails
The company says it will make the covert safeguard preventing model distillation as visible as other safety measures.
The company says it will make the covert safeguard preventing model distillation as visible as other safety measures.
Anthropic has apologized for stealthily throttling its new AI model, Claude Fable 5, with hidden guardrails that undermine both researchers and rivals using it to develop competing systems. The company says it is reversing course and will be more transparent about when the restrictions kick in, even if that means Fable refuses more queries.
Fable is the first widely available model in Anthropic’s Mythos class of AI systems, a group the company has spent months warning are too dangerous for public release. Anthropic says it has addressed some of those risks by launching Fable with safeguards that prevent it from responding to certain “high-risk” queries. One of the areas Anthropic said it would restrict Fable’s responses is distillation, a technique for training smaller AI models using the outputs of larger ones.
In Fable’s system card — a public document AI developers release to explain how a system works — Anthropic said it would handle queries it believed were distillation attempts by altering and degrading the model’s answers directly. Users would not be notified that they had triggered the safety measure or informed that the responses had been changed.
Anthropic said it is now changing its approach to distillation: Queries will now fall back to Claude Opus 4.8, Anthropic’s previous flagship model, the company said in a post on X. Anthropic will prominently tell users too: “You will see this every time it happens.”
This is similar to how Fable handles queries in other high-risk areas. When safety features are triggered in areas like biology, chemistry, and cybersecurity, queries are routed through Opus 4.8 unless they are blocked outright under the company’s broader safety rules, such as those covering drugs, weapons, or other prohibited content. In some cases, notably biology, the safeguards have been calibrated so broadly that Fable is practically unusable for even basic queries, something Anthropic spokesperson Paruul Maheshwary acknowledged in a comment to The Verge.
“Visible safeguards can be probed, so they have to be robust, which takes time to get right,” Anthropic wrote on X. “Invisible safeguards can be targeted more narrowly, allowing us to ship quickly with very few false positives. We went with 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.”