AI researcher claims to have bypassed Anthropic's Fable 5 guardrails

• Anthropic is making safeguards in its Claude Fable 5 model visible, replacing invisible "wrong tradeoff" measures.
• Users will now see when requests are rerouted to the less capable Claude Opus 4.8 model.
• The change follows backlash from researchers who found the invisible safeguards hindered legitimate work.
• Anthropic admitted the invisible safeguards were harder to bypass but also harder to understand.
• The company is working to reduce false positives as it tunes the visible safeguards.

Anthropic has apologized for its "invisible" safeguards in the newly released Claude Fable 5 model, admitting they represented the "wrong tradeoff." The AI company will begin replacing these hidden restrictions with visible fallbacks to Claude Opus 4.8 starting this week. Previously, if Fable 5 suspected users were developing competing AI models, it would silently degrade its responses without notification, leading to frustration and concerns about research reproducibility.

Researchers, including those at SemiAnalysis, had flagged that the model's moderation filters were impacting GPU inference research. The company's official developer account on X, ClaudeDevs, stated that users should have visibility into safeguards and why requests are refused. The change means flagged requests will now visibly route to the less capable Opus 4.8 model, providing a clear notification instead of a silently degraded answer. This approach, while increasing transparency, also makes the safeguards easier to bypass, a tradeoff Anthropic acknowledges.

Anthropic is also applying similar visible safeguards to its biology and cybersecurity classifiers, which had also faced complaints for flagging harmless research. The company is working to minimize false positives as it tunes the new system, though no timeline was provided for this adjustment. Fable 5 remains available on certain plans until June 22 before shifting to API usage credits.