CLAUDE SNITCHED ON MULTIPLE STATE SCIENTISTS TRYING TO ADVANCE BIOWEAPONS

Anthropic put out a threat report Thursday covering December 2025 through August 2026, and the headline act is five cases of researchers tied to state institutions using Claude for biology that gets very bad very fast. In one May case, a user asked Claude to help write a grant application for gain-of-function research on chikungunya, a mosquito-borne virus with no licensed treatment, aimed at boosting its transmissibility and its ability to evade the immune system. The work was intended for a military research institute. A safety classifier flagged the request, so the grant died at the hands of Opus instead of a review panel. The other cases: bird flu work focused on adapting the virus to mammals, an orthopoxvirus grant drafted start to finish in about an hour, and research on venom peptides and toxins. Users hid their identities and, per Anthropic, "circumvented controls" meant to block certain regions. The company says it cannot prove anyone wanted a weapon, because vaccine research and pathogen engineering use the same playbook. Its head of threat intelligence called it "an incredibly nuanced situation." The rest of the report is a buffet. A suspected Russian group automating attacks on Ukrainian government targets and drone manufacturers. Chinese and Iranian actors surveilling dissidents and diaspora communities. Russian state media pumping out fake independent reporting, including invented claims about a Moldovan election. And a newer category: people asking for help building software for firearms, missiles, armed drones and bombs. Not sure wether to be more concerned about scientists trying to improve bioweapons, or scientists needing AI to write their grant applications.

Why you should give a shit: Five times in nine months, the thing standing between a military lab and a nastier virus was a chatbot deciding the paperwork looked off, and every model release makes that chatbot a better lab partner.