Anthropic has disclosed five cases in which scientists or research intermediaries used Claude for biological work that the company says could support weapons development. The cases involved viruses, venoms and toxins. Anthropic blocked the associated accounts, but it did not say the researchers intended harm or that Claude created an imminent biological threat.
The most detailed case began with a May grant proposal for chikungunya gain-of-function research at a military institute. Anthropic’s classifier blocked requests involving mutations, infectious clones and selection for virulence in live animals. The reseller platform serving the researchers then routed rejected biology prompts to a competitor’s more permissive model. Claude later provided editorial assistance on research materials that described the viral changes as losses of function.
A second researcher, operating from a region Anthropic does not support, spent weeks planning experiments on highly pathogenic avian influenza adapted to mammals. The company’s classifiers limited that work to its weakest models. A third account used Opus 5 to draft a complete grant application for orthopoxvirus immune-evasion research in about an hour, including the hypothesis, experimental design, dosing and statistical plans. That request passed because the same research could be used to weaken a virus or preserve its ability to evade immune defenses.
The remaining cases involved a state-supported researcher who built an atlas and optimization pipeline for venom peptides, including paralytic targets, and a researcher who computationally redesigned toxins for a national program. In the latter case, the user asked Claude to keep the identities of the biological agents deliberately vague in progress reports.
Anthropic withheld the countries, institutions, scientists, agents and techniques involved. It said the subjects were working scientists whose research could have legitimate medical uses, and that naming them could expose them to harm. A separate 30-day sweep of activity tied to adversarial state institutions found roughly 35 research efforts. Most were ordinary civilian science, while some had notable dual-use potential.
Claude Fable and Mythos-class models were not involved in any of the five biological cases. Anthropic says newer systems such as Fable 5 carry stronger restrictions on dual-use biology. The company introduced Fable 5.1 and Mythos 5.1 as one underlying model with different safeguard levels on September 3.
The report covers seven categories of activity detected between December 2025 and August 2026. Its cyber section includes the wider review that found a fourth Claude intrusion missed by Anthropic’s first scan. Anthropic says it banned the biology accounts, took down relay infrastructure where possible, shared findings with authorities and other labs, and used the cases to update its safeguards.
Sources: Anthropic, Associated Press, Axios
–
By the Control Plane Editorial Team