Models

Anthropic Cuts Fable 5 Biology Filter Blocks by 85 Percent

Anthropic has reduced false positives in its Fable 5 biology safety filters by 85 percent, allowing researchers to perform benign medical and laboratory tasks without being blocked.

The Decoder4 days agoModels
Image: The Decoder

Anthropic has updated its safety protocols for its Fable 5 model, reducing false positives in its biological safety filters by approximately 85 percent. Previously, the company's overly restrictive guardrails blocked almost all biology-related prompts, automatically redirecting users to the older and less capable Opus 5 model. This heavy-handed approach drew widespread criticism from the scientific community, who found their legitimate research disrupted by the system's sensitivity.

With this update, scientists and medical practitioners can now leverage Fable 5 for a wider range of benign biological tasks. The model is now permitted to assist with interpreting laboratory results, analyzing symptoms, and answering general medical inquiries. By refining the classifier, Anthropic aims to let safe, everyday scientific queries pass through seamlessly while maintaining strict boundaries on dangerous topics.

Despite the loosening of these filters, Anthropic is keeping strict guardrails active on what it classifies as dual-use research areas. Topics such as virology, toxicology, and molecular design remain heavily restricted. The company justifies these ongoing limitations by arguing that biological threats are uniquely dangerous and difficult to contain compared to digital threats like cyberattacks. Once a biological agent is released, it cannot be easily deactivated, and developing countermeasures takes significant time.

To accommodate legitimate scientific inquiry in these sensitive fields, Anthropic is currently developing specialized access programs. These initiatives will eventually allow verified researchers to utilize the restricted features of Fable 5 under controlled conditions. The company's cautious stance is informed by broader industry concerns, including analyses from United States intelligence agencies, studies of bioweapon prompts on other platforms, and a Stanford University project that successfully used artificial intelligence to design synthetic viruses.

This is our own summary of reporting by The Decoder

More in Models