이미지: AI 생성 — METAL LAB
Summary
- Anthropic overhauled Fable 5's biology safety classifier, cutting false-positive-driven fallbacks by about 85%.
- General biology queries for health and educational purposes are now more broadly allowed, but dual-use areas like virology and toxicology continue to be blocked.
- When Fable 5 is blocked, requests fall back to Opus 5, while a separate trusted access pathway for professional researchers is under development.
Redesigning the Classifier to Cut False Positives
Anthropic is an AI safety research company focused on developing reliable, interpretable, and controllable AI systems. The update it announced is the result of a full rewrite of Fable 5's biology safety classifier. The company rewrote the classifier's rule set — its "constitution" — rebuilt the training data based on it, and then retrained the classifier. The process also involved gathering input from a range of internal and external expert groups.
The redesigned classifier is built to better distinguish between benign queries and dual-use queries, capturing subtler differences than before. As a result, biology-related fallbacks are expected to drop by about 85% across the product lineup. Specifically, Anthropic projects total fallback reductions of about 67% on Claude.ai, about 55% on Cowork, about 17% on Claude Code, and about 7% on Claude Platform.
What Is a Fallback
When Fable 5's safety classifier detects a risky or potentially dual-use biology query, it automatically routes the request to Opus 5 instead. Opus 5 does not have the same level of biological capability as Fable 5, which limits the assistance available to a malicious user. This rerouting is what users experience as a "fallback."
Anthropic explained that when Fable 5 launched, it deliberately applied a very broad classifier that blocked nearly all biology-related queries. This was a decision made to proactively manage the risk of biology misuse while still allowing the model to serve users normally in other domains. Rather than delaying the model's full launch until the classifier could be fine-tuned to a sufficient degree, Anthropic chose to cast a wide net first, release the model, and then refine the system in stages.
Why Biology Gets Special Treatment
Anthropic applies especially strong safeguards to the biology domain because distinguishing dual-use potential in this area is particularly difficult. Developing a treatment for a disease sometimes requires first generating the very hazardous compound that causes it. Live vaccines are a prime example, requiring scientists to cultivate the pathogen they are trying to prevent. Similarly, when researchers developed captopril, a hypertension medication, they had to isolate a toxic component of snake venom that sharply lowers blood pressure.
According to Anthropic's capability evaluations, Fable 5 has reached a level in some advanced biology tasks that surpasses human experts, and the company judged that this could provide malicious actors with capabilities otherwise unavailable elsewhere.
What Has Changed Now and What Remains Blocked
With this update, fallbacks are expected to drop significantly for everyday health and educational queries — such as interpreting lab test results, understanding symptoms, or general biology learning. Medical professionals will also be able to receive more support from Fable 5 on clinically related queries.
However, specialized research areas with dual-use potential — including virology, toxicology, and molecular design — still fall back to Opus 5, and Fable 5 cannot yet be used for professional biology research or drug development. Anthropic said it is separately developing a trusted access pathway to close this gap. The specific mechanics or timeline for this pathway have not yet been disclosed.
