Anthropic publishes report: five Claude bioweapons attempts blocked

Anthropic has put out a new misuse report saying it blocked five attempts in 2026 to use Claude for dual-use biology work that could feed into weapons development. The company says the cases involved users allegedly trying to get around safeguards and conceal what they were doing, and that some of the access was linked to banned regions, including Russia, China, and Iran. The examples are specific and ugly: weeks of avian-flu planning, a grant proposal for gain-of-function chikungungya meant to make it more transmissible and more dangerous, plus work touching orthopoxviruses, venoms, and toxins.

ChatGPT Download

Anthropic also says you can’t read harmful intent straight from prompts alone, because the same scientific knowledge can be used for vaccines and disease surveillance. Still, it banned the accounts, said the avian-flu activity never got beyond weaker models, used the incidents to toughen up Claude Fable 5, and says it published the report in hopes of pushing both companies and governments to tighten safeguards faster. If you follow AI safety, this one’s worth your time. Researchers have been warning that stronger chatbots could lower the knowledge barrier for biological threats in the wrong hands, even if lab access, materials, tacit know-how, and safe handling still matter a lot; Jennifer Doudna has said models by themselves aren’t enough; and a 2025 White House order still leaves oversight scattered just as the argument over AI moves past homework, coding, and search.

You can read the report on Anthropic’s site.

Claude Download