Anthropic's bioweapon filters were off for eleven months — across 133 million conversations
In its second company-wide Risk Report, published on 14 August, Anthropic disclosed that the blocking classifiers meant to stop its models helping anyone build a biological weapon were inactive on all human-feedback vendor traffic from May 2025 to April 2026. That covers roughly 133 million exchanges with about 50,000 external contractors, people vetted only by the outside vendors that hired them. Among the models exposed during that window was Mythos Preview, one of Anthropic's most capable, which sat behind the gap for around two weeks. The company says its internal investigation found no evidence of actual misuse, that no customers were affected, and that it has since tightened contractor requirements and closed the gap.
Why it mattersAnthropic's public case for its own safety framework leans heavily on biological weapons as the risk that justifies it — and the company did not notice its main filter for that risk was switched off for nearly a year. Safety promises rest on plumbing that can quietly break without anyone raising an alarm. It also shows where the raw, unfiltered model actually sits: not with paying customers, but with tens of thousands of contract workers screened by someone else.
Read in the app — free, in 9 languages
Related stories
Anthropic expects to raise as much as SpaceX did — the largest listing in history — and could file publicly this month2026-08-21Nvidia is paying $6 billion for the machine that builds a rival's models — and hiring 109 of the people who ran it
2026-08-21The doctors and lawyers who grade AI answers are now a half-billion-dollar business — and this one grew fivefold in eight months
2026-08-21Anthropic still leads on business AI spending. For the first time since May, the gap is closing rather than widening.
2026-08-21Alibaba's AI business grew at its fastest pace in five and a half years. Its profit fell about 75 percent.
2026-08-20