Anthropic Discloses Year-Long Gap in Biosecurity Filtering
Anthropic revealed in its August 2026 risk report that its biosecurity blocking classifiers — which stop models from giving out dangerous chemical and biological weapon information — were not applied to external contractor traffic from May 2025 to April 2026. During that window, roughly 50,000 contractors ran about 133 million chats with the models, with their trustworthiness vetted only by the contractors themselves, whose procedures Anthropic admits were often insufficient. An internal investigation found no evidence of abuse, and the company has since tightened contractor requirements.
The redacted report rates Anthropic’s catastrophic risks as low across all categories, but is noticeably more cautious than its predecessor — most estimates were raised or hedged with greater uncertainty, citing recent incidents and early signs of accelerating AI progress.
Related: Dario Amodei Proposes Sweeping AI Policy Framework Across Five Fronts, Anthropic’s Claude Surpasses PhD Experts in Bioinformatics with BioMysteryBench