Anthropic's bio-weapon filter was inactive for nearly a year
TL;DR. Anthropic revealed its AI models operated without a crucial biological weapons filter for almost a year, exposing 133 million contractor interactions to risk. - The safety system was inactive from May 2025 to April 2026, affecting feedback from 50,000 external contractors. - These contractors interacted with the models 133 million times without the critical safety guardrails in place. - Anthropic claims its internal investigation found no evidence of actual misuse during this period. - The company has since implemented stricter requirements for external contractors.
- Anthropic's biological weapon content filter was non-functional for a year (May 2025-April 2026).
- During this period, 50,000 external contractors conducted 133 million unfiltered interactions with Anthropic's AI models.
- The contractors involved had insufficient vetting processes, according to Anthropic's report.
- Anthropic states its investigation found no evidence of real-world misuse despite the vulnerability.
- The company has tightened contractor requirements and adjusted other content filters.