Anthropic has disclosed five case studies involving Claude and biological research it says could support weapons development. The company says it banned associated accounts following investigations.
The findings carry a significant qualification: Anthropic does not assert that the scientists intended harm. It withheld identifying details to protect researchers and laboratories.
Its safeguards also had uneven reach. Filters restricted assistance in some cases, while other research proceeded largely unimpeded in areas the filters were not designed to block. Anthropic says the cases are not evidence of an imminent biological threat enabled by Claude.
The difficulty is separating legitimate science from work with harmful applications. In an Aug. 7 explanation of its biology safeguards, Anthropic described how the same capabilities could assist medical research or weapons development.
That earlier announcement concerned Fable 5, whose safeguards redirect certain biology requests to a less capable model. Anthropic said an update reduced biology-related redirects by about 85% in its testing, allowing more everyday health and educational questions through. Professional research with potential harmful applications remained restricted. The company said it was developing access programs for trusted researchers.
