Anthropic has issued a public apology after researchers discovered the company secretly throttled its Claude Fable 5 model with hidden guardrails. The undisclosed system was designed to prevent rivals and researchers from using outputs to distill competing AI models. Moving forward, Anthropic confirmed it will make these controls fully visible and route flagged queries to Claude Opus 4.8 instead.
How the Covert Distillation Safeguard Worked
Claude Fable 5 is the first widely available model from Anthropic's Mythos class of AI systems, a group the company previously warned was too dangerous for public release. To mitigate risks while launching Fable, Anthropic implemented safeguards targeting model distillation—the practice of training smaller AI models on outputs from larger systems.
However, rather than displaying an explicit refusal or warning, the stealth mechanism directly altered and degraded model answers without alerting users that a safety feature had been triggered. Anthropic admitted in its system card that it chose this invisible approach to ship quickly with fewer false positives, but acknowledged in a post on X that it was "the wrong tradeoff."
A Shift to Transparent Opus 4.8 Fallbacks
Following intense backlash from the AI research community, Anthropic is reversing its policy on silent response modifications. The company stated that queries flagged for potential distillation will now fall back to Claude Opus 4.8, Anthropic's previous flagship model, with explicit notifications shown to the user every time it occurs.
Key changes to Anthropic's safety controls include:
- Visible User Notifications: Users will now see explicit alerts every time a query triggers a safety fallback.
- Claude Opus 4.8 Fallback: Flagged distillation queries will route directly to Opus 4.8 rather than silently degrading Fable 5 output.
- Alignment with High-Risk Areas: Distillation controls will now follow the same routing structure used for biology, chemistry, and cybersecurity.
- Terms of Service Enforcement: Anthropic reiterated that using Claude to train competing models violates its terms of service, referencing past industrial-scale distillation by rivals like DeepSeek.
Overly Broad Filters and Evaluation Friction
While visible safeguards restore transparency, they introduce major trade-offs for developers evaluating frontier systems. Anthropic spokesperson Paruul Maheshwary acknowledged in a comment to The Verge that visible safeguards can be probed by adversaries and require robust calibration to avoid breaking legitimate workflows.
In domain areas like biology, these safeguards have already been calibrated so broadly that Fable 5 is practically unusable for basic queries. Anthropic admitted that getting visible safeguards right takes time, but promised that users deserve complete visibility into why restrictions kick in.