Anthropic Reverses Secret Plan to Thwart Rivals Using Its AI Model

Anthropic has scrapped a policy that would have quietly sabotaged researchers trying to use its latest AI model, Claude Fable 5, to build competing systems. The company reversed course after a wave of criticism from the AI community, admitting it made the wrong call.

Earlier this week, Anthropic released Fable 5 with safety guardrails to prevent misuse. Some restrictions were straightforward—users asking about cybersecurity, biology, or chemistry would be redirected to a less capable model to reduce risks of cyberattacks or bioweapons. But for researchers working on frontier AI development, the company took a different approach: it would secretly degrade the model’s performance, making it harder to train rival AI systems. The move violated Anthropic’s own terms of service, which ban using Claude to develop competing models.

“We made the wrong tradeoff and we apologize,” Anthropic said in a statement. The company now says the safeguards will be visible to users. If it suspects someone is using Claude to build a highly capable AI, it will either refuse the request or redirect them to a weaker model—and tell them why.

The policy shift follows backlash from researchers who argued that secret degradation would stifle collaboration and concentrate AI research in a handful of labs. “It felt like Anthropic was saying, ‘We don’t trust anybody else to do AI research. We are the only ones who have to do it,’” said Will Brown, research lead at open-source AI startup Prime Intellect. “It feels a bit like they’re starting to pull the ladder up behind them.”

Anthropic said it implemented the measures because Claude has become increasingly effective at accelerating AI research, and it worries that AI could outpace society’s ability to adapt. The company also cited concerns about foreign adversaries using its models to optimize chips or develop dangerous capabilities. By making the safeguards visible, Anthropic says it needs to cast a wider net, meaning more benign requests may trigger restrictions. It’s working to improve precision.

For now, the reversal offers relief to developers who rely on Claude for open-source research. But the episode has left many questioning Anthropic’s commitment to transparency.