Anthropic has updated the biology safeguards on its Claude Fable 5 AI model, substantially reducing false positives and enabling the model to handle a broader range of biology-related tasks. According to the company, the update has reduced biology-related fallbacks by approximately 85 per cent across its product surfaces.
Previously, when Fable 5 detected a biology-related query, it could switch to a less capable model. Following the safeguards update, users should experience fewer such fallbacks, particularly when asking everyday health and educational questions. These include interpreting laboratory results, understanding symptoms and learning about biology in educational contexts.
The update is also intended to expand support for healthcare professionals undertaking clinical tasks. Anthropic said the changes are designed to make Fable 5 more useful for legitimate biology applications while maintaining safeguards around higher-risk capabilities.
Despite the reduction in restrictions, Fable 5 will continue to fall back to Opus 5 for requests considered dual-use, including areas such as virology, toxicology and molecular design. As a result, the model is not yet intended for unrestricted professional biology research or drug development.
Anthropic said it is working towards expanding access to frontier biology capabilities through trusted access pathways, alongside improvements to its safety classifiers.
The company said the decision to introduce stronger biology safeguards at Fable 5’s launch was driven by the model’s growing capabilities in biological science. According to Anthropic, Fable 5 can outperform experts on some highly complex biological tasks and provide operational support on others. While these capabilities could accelerate scientific research and medical treatment development, the company noted that the same capabilities could potentially be misused for harmful purposes, including biological weapons development.
A key challenge, Anthropic said, is distinguishing beneficial biological research from potentially harmful activity. Some legitimate medical research can involve the production or study of dangerous biological or chemical substances. The company cited vaccine development and the historical development of the hypertension drug captopril, which involved research into toxic components of snake venom, as examples of the complexities involved in drawing a clear line between beneficial and harmful applications.
Anthropic also pointed to its capability assessments, which found that Fable 5 could provide significant “uplift” to malicious actors in certain biological tasks. The company said sophisticated actors could potentially disguise harmful requests as legitimate scientific research, making automated safety screening particularly challenging.
The company therefore initially chose to block almost all biology queries when launching Fable 5, accepting a high rate of false positives in exchange for reducing potential misuse risks. The latest update represents a shift towards more targeted safeguards, allowing more legitimate biology-related queries to be handled by the frontier model while retaining restrictions around higher-risk dual-use capabilities.
Anthropic said its longer-term objective is to provide researchers and biologists with greater access to advanced AI capabilities while developing trusted mechanisms that can help manage the risks associated with frontier AI in biology.
