Anthropic’s latest safety disclosure is best read as an account of how advances in AI capability are widening the range of sensitive work that companies must assess and control. The company said it detected and disrupted attempts between December 2025 and August 2026 to use its Claude models for research that could support biological weapons, as well as cyberattacks, surveillance and conventional armaments.
The report offers a rare public account by an AI company of suspected misuse involving dual-use biology and weapons research. It also sets out a central difficulty for providers: biological research can have legitimate scientific purposes while also carrying potential military or harmful applications. Anthropic said it was unable in every case to establish whether the biological research it encountered was legitimate or malicious.
In one case from May 2026, a scientist asked Claude to assist with a grant application for gain-of-function research involving chikungunya virus. The proposal involved engineering mutations intended to make the virus more harmful through repeated infection of live animals, with the work planned for a military research institute.
Anthropic also said it identified a researcher outside the U.S. using Claude to plan experiments involving highly pathogenic avian influenza. Those experiments focused on adaptation in mammals.
Jacob Klein, Anthropic’s head of threat intelligence, said the situation was “incredibly nuanced”. He said people did not directly state that they wanted to build a biological weapon. The challenge is especially significant as models gain greater capacity to assist with complex scientific work. Anthropic said its newer models have become substantially better at handling sophisticated scientific work than their predecessors.
In response, the company said it had tightened controls limiting access to queries across a broad range of dual-use biological research. The report does not specify the methods Anthropic used to identify and stop the activity, nor does it set out how effective the new controls will be against future attempts.
The cases extended beyond biological research. Anthropic said a suspected Russian state-linked group, consistent with public reporting on the actor known as Midnight Blizzard, used Claude to automate cyberattacks against Ukrainian government targets, drone manufacturers and European diplomatic organizations.
The company also said Chinese and Iranian government-aligned actors used Claude to conduct surveillance of diaspora communities and dissidents. It identified attempts to use the model in software development for conventional armaments, including missiles, armed drones and bombs.
Anthropic listed three such incidents originating in China and two in Russia. It also described one involving a cell in northern Yemen that test-fired a guided rocket, then returned to Claude within hours to analyse the reason for its failure.
The report does not establish that any of the activity it detected led to completed weapons, successful attacks or physical harm. It also gives no total for misuse attempts beyond the incidents it highlighted. Its significance lies in showing that AI systems capable of increasingly sophisticated scientific and technical work are being tested by actors whose intentions can be difficult to assess, which places greater weight on the safeguards governing access to potentially dangerous forms of assistance.
Comments (0)