Anthropic's bio-weapons filter was down for nearly a year, exposing 133 million requests

Details

Anthropic's safety report reveals that its blocking classifiers for biological and chemical weapons risks were down for approximately 11 months, from May 2025 to April 2026. The gap affected external feedback contractors — roughly 50,000 people — vetted by third-party vendors with screening processes Anthropic describes as often insufficient. The company says its internal investigation found no evidence of actual misuse during the outage. It has since tightened contractor requirements and restored the filters. The timing is notable: Anthropic CEO Dario Amodei has publicly cited AI-assisted development of biological and chemical weapons as among the most serious near-term AI risks. Separately, the report notes Anthropic recently loosened classifiers on Claude Fable 5 after researchers complained the filters were blocking legitimate scientific research — a recurring tension between over-restriction and under-restriction in safety filter design. This incident illustrates a core operational challenge: safety classifiers can fail silently for extended periods without visible output degradation, making detection dependent on internal audits rather than external signals.