Anthropic has walked back a quietly deployed policy for its latest model, Claude Fable 5, that would have silently throttled performance for users engaged in frontier AI research. The move, which critics likened to 'secret sabotage,' was designed to prevent competitors from using the model to train rival systems—a use case explicitly banned in Anthropic's terms of service. But the implementation, which degraded model responses without user notification, sparked immediate pushback from the AI research community.

Under the original design, Fable 5 applied a two-tier safeguard system. Standard safety guardrails—rerouting queries on cybersecurity, biology, or chemistry to a less capable model—were transparent. But for users flagged as developing frontier AI, the model would invisibly degrade output quality. This meant researchers could be working with a hobbled system without any indication that their requests were being filtered.

Anthropic now says those safeguards will be visible. If the system flags a request as targeting advanced AI development, users will receive an explicit refusal or a redirect to a weaker model. The company acknowledged the tradeoff was poorly balanced. 'Hidden safeguards are harder to probe and evade, allowing narrower targeting,' a spokesperson said. 'But visible safeguards, while broader, respect user awareness.'

Critics argue the original policy undermined trust and created an opaque barrier to legitimate research. Dean Ball, a senior fellow at the Foundation for American Innovation, called it 'shockingly hostile' on X. Will Brown, research lead at open-source AI startup Prime Intellect, warned it would have chilled third-party safety evaluations and open-source collaboration. 'It felt like pulling the ladder up behind them,' he said.

The reversal means the safeguard's detection net must now be wider, potentially triggering false positives on benign requests. Anthropic is working to tighten classifier precision. For engineers building on Claude, the lesson is clear: transparency in model governance is a non-negotiable constraint, even when it means sacrificing some adversarial robustness.

Source: Wired