Anthropic blocked five attempts to use Claude for bioweapons and admits it can't tell whether they were science or threat

By Zendoric · September 10, 2026.

Anthropic said on Thursday that it has detected and blocked five cases this year of scientists using its AI models — the Claude family — in ways that could contribute to the development of biological weapons. The company details this in a new report on misuse of its systems which, by its own account, is not representative of typical usage but is serious enough to publish.

According to Anthropic, those responsible circumvented the controls that block access from unauthorized regions and resorted to other tactics to disguise the real purpose of their research. The company says it has suspended the accounts involved and has fed what it learned into its safety, compliance and threat-detection systems «to better prevent, detect and disrupt this kind of activity in the future», the report states.

One of the cases, dated May, describes a scientist who asked Claude for help drafting a grant application to research the chikungunya virus, which is transmitted by mosquitoes. Legitimate research on this pathogen can lead to vaccines; in this case, according to Anthropic, the goal was to design mutations that would make the virus more harmful after repeated infections in live animals. One detail raised suspicions: although the application appeared to be signed by civilian researchers, the research was going to be carried out at a military institute. Anthropic has not disclosed the country, the institution or the specific pathogen in any of the five cases.

The report also covers attempted misuse in cyberattack operations, mass surveillance and conventional weapons development; other outlets that have had access to the document additionally mention cases linked to avian flu and to drone swarm systems. Even so, the company admits something uncomfortable: it cannot always determine whether it was looking at legitimate science or a real threat, because biology is, by its very nature, dual-use.

«Sophisticated actors are aware that we try to detect dangerous uses of our models, and they exploit that dual-use nature of biology to maintain a kind of plausible deniability about their research», the report notes. «Overt malicious intent is, in fact, often evidence that an actor is not very sophisticated. The most sophisticated actors can hide their intent.»

Andrew Weber, of the Council on Strategic Risks, described the cases as «chilling» examples of state-backed actors exploiting the capabilities of the most advanced AI models, as he told The New York Times. The report also comes at a time when several Anthropic employees are publicly warning about AI's safety and existential risks, after a former researcher at the company warned that AI companies are heading «straight for self-improving superintelligence, gambling with our lives», as reported by HuffPost.

Our read: this report lays bare, more clearly than any other so far, the tension that defines AI's near-term biological risk. The same capability that lets a model help design a vaccine serves, in the wrong hands, to make a virus more lethal; the model does not distinguish intent, it merely executes plausible instructions. That forces Anthropic — and presumably the rest of the frontier labs, which face the same dilemma — to monitor behavioral patterns (institutional affiliation, access geography, inconsistencies between what is declared and what is real) instead of trying to read intentions, because intent, when the actor is competent, is precisely what gets hidden.

It is also worth being honest about the limits of this transparency exercise: it is Anthropic itself that decides which cases to publish, how to describe them and what to leave out, with no independent verification. That does not invalidate the report, but it is a reminder that self-regulation by AI companies is no substitute for external oversight; at best, it is a necessary complement while that oversight is being built.

In the long run, this same technology underpins the thesis that AI can help eradicate disease and extend healthy life: chikungunya, avian flu and dozens of dual-use pathogens are also targets of vaccines and biomedical research accelerated by these very models. What this report leaves open is not whether AI will accelerate biology — it already does, in both directions — but whether the governance being built today will be fast and precise enough to tip that balance toward the cure before the weapon.