Anthropic Says It Blocked AI Users Linked to China, Russia and Yemen From Bioweapon and Missile Research

Anthropic published a threat intelligence report on Thursday, September 10, 2026, describing eight months of attempts by state-linked actors and researchers to misuse its Claude AI models, according to the company's own findings as reported by CNN, NPR, Fox News, Breitbart and Insurance Business Magazine.

The report covers activity from December 2025 through August 2026. Anthropic says it looked at 30 days of usage and flagged about 35 "distinct research efforts" with potentially concerning biological activity, according to CNN. Five of those became formal case studies.

The Bioweapon Cases

One case from May involved a scientist requesting help writing a grant application for chikungunya research, a mosquito-borne virus that causes severe joint pain. According to Fox News and Breitbart, the researcher was working to engineer mutations that would make the virus more transmissible and better at evading the immune system, then pass it repeatedly through live animals, a method called gain-of-function research. Anthropic said the work appeared linked to a military research institute.

Jacob Klein, Anthropic's head of threat intelligence, told Breitbart the case wasn't an obvious comic-book villain scenario. "It's an incredibly nuanced situation," he said. "What we don't know is if the research was meant to be weaponized. But a military institution doing gain-of-function research is concerning."

Other cases involved bird flu research focused on mammalian adaptation and severity, work on orthopoxviruses (the family that includes smallpox and mpox), and redesigning venom toxins, according to NPR and CNN. NPR reported one toxin case was tied to what Anthropic described as a "national program."

Anthropic did not name any country, institution, or specific pathogen strain in any of the five cases. The company told CNN it could not determine whether the researchers "intended harm" or were doing legitimate science, but banned the accounts anyway.

Missiles, Drones and the Houthi Connection

Breitbart's coverage highlighted a category not previously flagged by Anthropic: attempts to use Claude for conventional weapons development, including missiles, drones and firearms. The report cited three cases in China, two in Russia and one in Yemen.

Anthropic didn't name the Yemen group, but Breitbart noted the report's description points to the Iran-backed Houthi militia. That case reportedly touched multiple missile programs, including a multi-stage ballistic missile with a stated range over 2,000 kilometers and a system with a hypersonic glide vehicle variant. Anthropic said the actors used Claude Code to write guidance, navigation and control software in place of human engineers.

Espionage, Propaganda and Cyber Operations

Government-linked actors from China and Iran reportedly used Claude to help surveil dissident and diaspora communities, per Breitbart. Russian state media allegedly used the model to produce propaganda dressed up as independent journalism, including fabricated claims tied to Moldova's election.

Insurance Business Magazine, citing the BBC, added detail on the cyber side: a group consistent with Russia-linked Midnight Blizzard built AI workflows that detected when malware got flagged and rewrote the code to evade detection, and compromised hotel Wi-Fi at three hospitality vendors via DNS hijacking to target government and defense officials. ShinyHunters affiliates reportedly moved from initial access to bulk data theft in two to three hours at one target, pulling over a terabyte of data from a tech provider and tens of millions of airline passenger records.

Anthropic's own conclusion, quoted by Insurance Business: "The main distinguishing feature between these classes of actors is no longer sophistication but intent."

What's Unproven, and Who's Grading Their Own Homework

This entire report is Anthropic policing itself and telling the public how well it did. No government agency has announced an independent investigation, filed charges, or issued sanctions tied to any of these cases. Anthropic withheld every country, institution and pathogen name that would let outside researchers or regulators verify the claims.

That's a legitimate concern for anyone skeptical of self-reported corporate security theater: a company grading its own safeguards has an incentive to make the disruptions sound serious and the fixes sound effective. Anthropic's repeated admission that it "could not always determine" intent cuts both ways. It's honest about the uncertainty, but it also means none of these cases has been confirmed as an actual bioweapons or missile program by an independent body.

NPR's report carries a disclosure: Anthropic is a financial supporter of NPR. NPR states it doesn't accept money in exchange for coverage, but the relationship is relevant context for readers weighing how the story got framed.

Fox News also tied the report to a broader, separate controversy, quoting former Anthropic and OpenAI researcher Jacob Coxon, who resigned warning of an "AI takeover," and Senator Bernie Sanders' call to "hit the pause button" on AI development. Those are real quotes about real AI safety anxieties, but they're a different debate from what Anthropic's own report actually documents, and Fox's framing blurs the two together.

Insurance Business Magazine flagged the practical fallout: Moody's listed cyber as a top corporate risk this week, warning AI is compressing attack timelines, while Lockton data shows average cyber insurance premiums fell roughly 11% in 2025 even as incident frequency climbed. That gap between falling premiums and rising AI-enabled attack capability is the one concrete, measurable consequence in this story, and it's one underwriters will have to price in without knowing which countries or labs were actually involved.