Anthropic's Safety Filters Failed for Nearly a Year

Anthropic released a safety report on August 14, revealing that some of its safety filters for blocking chemical and biological weapon-related dangerous requests had failed for a long time. Between May 2025 and April 2026, about 50,000 external contractors generated 133 million conversations that went unfiltered. The company said an internal investigation found no evidence of actual misuse.

The filters are part of Anthropic's multi-layer safety measures for CBRN risks, analyzing user inputs and model outputs to block dangerous information. The issue stemmed from insufficient vetting of external contractors, and Anthropic has tightened its screening requirements.

Earlier, Anthropic had enabled ASL-3 protections with Claude Opus 4 and continues to supplement with red-team testing and bug bounty programs. The company also acknowledged that overly strict filters could hinder legitimate research, so many biology and chemistry requests are now routed to Claude Opus 4.8.