# Anthropic 放宽 Fable 5 生物学限制，但保留病毒学与毒理学防护

- 来源：The Decoder：AI News（RSS）
- 作者：Matthias Bastian
- 发布时间：2026-08-08 01:35
- AIHOT 分数：48
- AIHOT 链接：https://aihot.virxact.com/items/cmsj8ffib02kyroo5sa8d9vwp
- 原文链接：https://the-decoder.com/anthropic-loosens-fable-5s-biology-restrictions-but-keeps-the-guardrails-on-for-virology-and-toxicology

## AI 摘要

Anthropic 将 Fable 5 生物学安全过滤器的误报率降低约 85%，此前几乎所有生物学查询都被拦截并转至能力较弱的 Opus 5。用户现可用 Fable 5 处理解读化验结果、理解症状、回答医学问题等更多任务。病毒学、毒理学和分子设计等“双重用途”主题仍受限，Anthropic 正构建访问计划供研究人员未来使用受限功能。

## 正文

Anthropic has cut false positives in its biology safety filters for Fable 5 by about 85 percent. Previously, nearly all biology queries were blocked and rerouted to the less capable Opus 5, drawing sharp criticism from scientists. Users can now handle far more biology tasks with Fable 5, such as interpreting lab results, understanding symptoms, and answering medical questions.

The classifier now lets most benign biology queries through while still blocking dual-use research topics. | Image: Anthropic

Restrictions remain for "dual-use" topics like virology, toxicology, and molecular design, where Anthropic says Fable 5 could give bad actors capabilities unavailable elsewhere. The company is also building access programs so researchers can eventually use these restricted features.

Anthropic's rationale: biological risks may be harder to contain than cyberattacks. A released virus can't be shut down, and countermeasures take time. An analysis by U.S. intelligence agencies, a review of ChatGPT chats requesting bioweapon instructions, and joint warnings from researchers confirm the threat could be real, so this isn't just the usual Anthropic Angst.