Anthropic Self-Reports 11-Month Bioweapon Classifier Gap Across 133 Million Contractor Exchanges
Anthropic's August 2026 Risk Report discloses that from May 2025 through April 2026, biological-harm classifiers were disabled and not logging in contractor feedback systems, affecting roughly 133 million exchanges with approximately 50,000 people. A retrospective review using Claude Sonnet 5 flagged 1,197 transcripts, but manual review of the 62 non-red-team cases found no clearly concerning misuse. The findings are Anthropic's own self-assessment; no regulator or independent auditor has reviewed the conclusions, and the report does not define the criteria used to grant classifier exemptions.