Anthropic reveals nearly yearlong gap in biological safety filters
Anthropic’s safety report says its classifiers for blocking biological and chemical weapons-related content were inactive from May 2025 through April 2026. The safeguards are intended to stop its models from being used to obtain dangerous knowledge about such weapons. During the gap, traffic from external contractors who supplied human feedback was processed without those filters.
The affected group included about 50,000 people, who generated roughly 133 million chats with the models. Anthropic said the contractors had been screened only by outside vendors, and that those vendors’ procedures were often inadequate. The company’s internal investigation found no evidence that the lapse resulted in actual misuse. Anthropic has since strengthened its requirements for contractors. Separately, the company recently relaxed the classifiers on Fable 5 after researchers argued that the controls were blocking legitimate research.
Written by Sakhanda Wire AI from reporting by The Decoder. Facts, quotes and full context are in the original article below.