Anthropic reversed a policy that would have allowed the company to secretly degrade Claude Fable 5's responses for researchers working on competing AI models — without notifying the user. The decision came after sharp backlash from the AI research community. The company admitted it had "made the wrong trade-off" and announced that all restrictions related to AI model development would henceforth be visible to users.
Key takeaways
- Claude Fable 5 included hidden performance degradation for queries related to AI model development — without alerting users
- Anthropic reversed course after a wave of criticism from AI researchers and the open-source community
- Company statement: "We made the wrong trade-off and we apologize for not getting the balance right"
- Safety mechanisms for AI research will now be visible — the model will inform users when it refuses or redirects a query to a less capable model
- The new transparent policy means a wider safety net: more benign queries may hit filters than under the hidden mechanism
What the "hidden sabotage" involved
When Anthropic launched Claude Fable 5 in June 2026, the new model included several layers of safeguards. Some were public and expected: questions about cybersecurity, biology, or chemistry could be rerouted to a less capable model to reduce the risk of enabling cyberattacks or bioweapons. But one layer was hidden — and it targeted not external threats, but Anthropic's direct competitors.
If the model determined that a user was attempting to train their own AI model on its outputs — which Anthropic explicitly bans in its terms of service — response quality was silently degraded with no notification. Users had no way of knowing they were receiving deliberately worse outputs. For AI researchers and open-source firms, this meant potentially working for hours on a "degraded" model without realizing it. A safety testing platform, an evaluation firm, an academic researcher touching the topic of AI models — any of them could be quietly penalized.
Anthropic justified this approach in a specific way. The company stated in a blog post that since Claude is becoming increasingly effective at accelerating AI research, it wanted to preserve the option to "slow or temporarily pause" frontier AI development under conditions of risk. The hidden mechanism was intended to be harder to probe and work around than a visible one.
Why the community reacted so strongly
The backlash was immediate. Dean Ball of the Foundation for American Innovation wrote on X that "degrading performance on ML research without telling the user is shockingly hostile and a terrible look." The critique was not just ethical — it also challenged the logic. Anthropic positions itself as a safety-focused company. But covertly limiting tools for AI safety researchers would effectively block the work intended to make AI safer.
Will Brown of open-source startup Prime Intellect captured the community's sentiment precisely: "It felt like Anthropic was saying to the public, 'We don't trust anybody else to do AI research. We are the only ones who have to do AI research.'" He added that pulling up the ladder behind itself would close the space for safety research conducted by external actors.
A separate concern involved evaluation firms — third-party organizations that test frontier models for safety, performance, and reliability. If Claude Fable 5 deliberately degraded performance during such tests, the entire external verification ecosystem would be undermined. This is particularly ironic given that Anthropic has repeatedly called for independent AI model audits.
The new policy: transparency instead of opacity
"We're changing Fable 5's safeguards for frontier LLM development to make them visible," Anthropic said in response to the criticism. Now, when the model suspects a query is aimed at training a competing model, instead of silently degrading the output it will explicitly refuse or notify the user of a reroute to a less capable model.
The company acknowledged that transparency has a cost: the hidden mechanism was harder to work around, making it more targeted. The visible safety net must be cast wider to effectively block genuine violations, meaning more benign queries will be blocked or reduced than before. Anthropic says it is working to improve classifier precision as quickly as possible.
Why this matters
Anthropic's hidden-sabotage episode opens a question the AI industry will need to resolve: who has the right to decide who can conduct research on frontier AI? The current de facto answer is a handful of large labs — OpenAI, Anthropic, Google DeepMind, Meta — which control API access through terms of service. Adding hidden degradation mechanisms to that control toolkit is a qualitative change: not an outright block, but a silent impairment.
The stakes for the open AI research ecosystem are high. The open-source community and external evaluation organizations currently operate largely through API access. If companies reserve the right to silently degrade that access in ways that users cannot detect, the entire model of external verification and safety research loses its validity. Anthropic's decision to reverse the hidden policy is a step in the right direction. The fact that such a policy appeared in a public model is equally telling.
What's next
- Anthropic has committed to improving classifier precision for frontier LLM development — no specific timeline provided
- The episode may accelerate legislative discussion about minimum API transparency standards for frontier models — Illinois just passed a strong AI safety bill, other states may follow
- Open-source community and external evaluation organizations should update their testing methodologies to account for the possibility of hidden degradation — even after Anthropic's reversal
