Anthropic’s internal filters for chemical and biological weapons risks were inactive for nearly a year on traffic from external feedback contractors, according to The Decoder, which cites a company safety report. The report says the affected classifiers were offline from May 2025 through April 2026 for that contractor workflow. The filters are described as systems designed to stop model users from extracting dangerous knowledge related to chemical or biological weapons. During the lapse, The Decoder reports, about 50,000 external contractors providing human feedback ran roughly 133 million chats with Anthropic models without those filters active. The contractor population matters because it was outside Anthropic’s direct employee base. According to The Decoder, Anthropic said the individuals were vetted by external vendors and that those screening processes were often insufficient. The company has since tightened contractor requirements, the report says. Anthropic also said its internal investigation found no evidence of actual misuse, according to The Decoder. That is an important limit on the story: the provided material supports a failure of a safety control across a large volume of interactions, not a finding that contractors used the models to develop chemical or biological weapons. The lapse is notable because Anthropic has publicly emphasized biological and chemical misuse as a major frontier-model risk. The Decoder frames the disclosure against that backdrop, noting that the company’s chief executive has described AI-assisted development of chemical and biological weapons as a bigger threat than cyberattacks. The same report also says Anthropic recently loosened classifiers on Fable 5 after researchers complained that the filters were too aggressive and blocked legitimate research. That detail points to the operating tension around these systems: classifiers must be restrictive enough to reduce misuse risk, but not so broad that they suppress benign scientific work. For now, the evidence in this cluster comes from one outlet summarizing Anthropic’s safety report. The core facts should be treated as Anthropic’s disclosed account as reported by The Decoder: a nearly year-long inactive filter path, roughly 133 million contractor interactions, no evidence of misuse found, and subsequent tightening of contractor requirements. Who benefits: Anthropic benefits if the disclosed remediation and tighter contractor requirements restore confidence in its feedback pipeline. Researchers may also benefit from more calibrated classifiers if the company can reduce overblocking without weakening high-risk safeguards. Who's exposed: Anthropic is exposed to scrutiny over how a safety control stayed inactive for nearly a year. External feedback programs are also exposed because vendor screening and contractor access are now central facts in the reported lapse.