September 11, 2026, (Inside AI) — Anthropic disclosed that it blocked multiple attempts this year by scientists in China and Russia seeking to use its Claude AI models for research that could aid biological weapons development. The findings appear in the company's third misuse report since March 2025.
The report details at least six conventional weapons cases tied to users in China, Russia, and Yemen. Those attempts involved code linked to firearms, missiles, armed drones, and bombs. Anthropic did not name the Yemeni group, but report details point to the Iran-backed Houthi movement.
Anthropic also found activity it believes was linked to government actors in China and Iran, including efforts to watch dissidents and expatriate communities. In a separate case, Russian state media used Claude to create online material designed to look like independent news, including false claims about an election in Moldova.
One biological case dates to May, when a scientist used Claude while preparing a grant application tied to chikungunya, a mosquito-borne virus that can cause severe pain and fever lasting months. The work involved gain-of-function research, where scientists alter an organism to study how changes affect it.
Anthropic said the scientist wanted to build mutations that would raise the virus's transmissibility and help it escape the immune system, testing this by passing it repeatedly through live animals. The company said the case raised more concern once it found signs the work was meant for a military research body. Anthropic did not name the scientist, the institution, the country, or the biological agents involved.
Anthropic said it could not always tell whether the research was meant for weapons work or legitimate medical study, such as vaccine development. The company said this is a hard line to draw in biology, where the same research can help people or cause harm. Because of this, Anthropic chose to act with caution and block the activity rather than wait for certainty.
The people behind the biological research cases took steps to hide the true purpose of their work and get around safety checks. Anthropic has since banned accounts linked to this activity. Tests last year on older models showed they could not give real help with dangerous biological research. Newer models can now handle far more complex scientific tasks, which led Anthropic to bring in tighter rules covering a wider range of biology-related queries.
Anthropic said its older models, such as Claude Opus 4 and Claude Sonnet 4.5, were not capable enough to meaningfully help a skilled user carry out dangerous biological research. That is no longer the case with its newest models, which is why stronger controls are now in place.
Why Public Disclosure Matters
Anthropic said it is sharing the report because it believes it has a duty to disclose harmful use of its services. The company said that as AI models grow more capable, the risks they pose will also grow unless AI firms and other groups working on safety take steps to reduce them.
None of the cases involved Anthropic's newest Claude Fable or Mythos-class models, apart from one case involving what the company called a large-scale, hidden effort to copy the abilities of one of its models into another system without permission.
Some experts say AI companies such as Anthropic and OpenAI are in a difficult spot, since they are left to decide what counts as safe or unsafe use without government oversight or public input. John Thickstun, an assistant professor of computer science at Cornell University, raised this concern.
Safety Gaps And Departures
The report came out two days after an Anthropic researcher, Jacob Coxon, announced he was leaving the company. Coxon said he was resigning over concern that Anthropic and rival firm OpenAI are moving too fast towards highly advanced AI systems. He said some of his colleagues now believe AI could pose a danger to human life before the end of the decade.
Anthropic said that in every case set out in its report, it managed to stop the harmful activity, used what it learned to strengthen its safety systems, and shared details with government bodies and industry partners.
