Anthropic says it blocked AI use that could support bioweapons work
Five case studies on biological weapons, six on conventional arms, and a catalogue running from hotel Wi-Fi scams to surveillance of dissidents.
Anthropic says it has blocked scientists who were using its AI in ways that could have supported the development of biological weapons, one finding in a threat report that also covers conventional arms design, surveillance of dissidents and a Russia-linked espionage campaign.
The company published the report on Thursday, covering misuse of its Claude Haiku, Sonnet and Opus models detected between December 2025 and August 2026. It said it did so because it believes it has "a responsibility to disclose malicious misuse of our services".
The five cases that matter most
The report sets out "five case studies of actors using our models in ways that could support biological weapons development", and Anthropic is direct about where that sits on its own list of concerns.
Biological misuse is "one of the most serious risks of frontier AI model", the company said, warning that without the correct safeguards such capabilities "could have catastrophic consequences".
It also explained why this is harder to police than it sounds.
"The same information that can be used to develop a biological weapon could also be used to develop", Anthropic said, "a vaccine or a cure for a disease".
That symmetry is the whole difficulty. A model cannot refuse dangerous knowledge without refusing the research that shares it, and the distinction lies in intent rather than in content.
Jacob Klein, Anthropic's head of threat intelligence, told the New York Times it was "an incredibly nuanced situation".
"You are not seeing someone in a comic book kind of way say, 'Hey, I want to build a biological weapon to kill everybody,'" he said.
Conventional weapons, and everything else
Alongside the biological cases, the report describes six instances where Claude was used "to develop software for conventional weapons, including firearms, missiles, armed drones, bombs, and other munitions" — and for "the targeting and control systems that operate them".
The rest of the catalogue is more mundane and, in volume, probably more consequential. Over eight months the cases ranged from fake dating apps and hotel Wi-Fi scams to surveillance tools built to identify dissidents. The report lists suspected state-sponsored groups, criminals, spyware vendors, state propaganda institutions and politically motivated individuals.
Claude was used in a Russia-linked cyber espionage campaign and by an Iranian propaganda institution. The hacking group ShinyHunters was among those named.
None of the misuse cases involved Claude Fable or the Mythos-class models, with a single exception involving distillation — the technique of training a smaller model on a larger one's output.
What the company did about it
Anthropic said it had incorporated its findings into its processes "to better prevent, detect, and disrupt these activities in the future", and that it had shared intelligence with authorities and industry partners where appropriate.
A threat report from the vendor whose product was misused is not an independent audit, and it is worth reading as both disclosure and positioning. The company is describing attacks it detected and stopped — the cases it did not detect would not appear.
The argument it lands in
The timing places this inside a larger dispute about the pace of AI development.
The report is the company's first this year, and it follows a warning from a top Anthropic safety researcher that AI is advancing so quickly that there is a greater than 10% chance it "could kill all humans" within the next decade.
On 6 September, OpenAI chief scientist Jakub Pachocki called for the industry to adopt "voluntary slowdowns" until safeguards are established.
"I am concerned no one is prepared for the consequences of a continued rapid rise in machine intelligence," Pachocki said, adding that OpenAI would continue building safeguards but that broader interventions were required.
That prompted an open letter to UK Prime Minister Andy Burnham calling for a multinational treaty on the safe development of AI, and for governments to work out what a treaty covering superintelligence should contain.
In the United States, the Democratic lawmaker Bernie Sanders has introduced legislation to ban AI superintelligence and temporarily pause advanced AI development.
"When scientists tell you there is a chance, a chance that it could have a cataclysmic impact on humanity," Sanders said on Thursday on BBC's Newsnight, "you've got be a moron not to say, slow it down."
Why this report is the awkward one
Extinction arguments are contested and unfalsifiable in the short term. This report is neither.
It describes specific people, over a specific eight months, using a commercially available model to work on biological and conventional weapons — and it comes from the company selling the model. The claim is not that something terrible might happen at some point. It is that it was already being attempted, repeatedly, and had to be stopped.
For legislators weighing whether AI risk is speculative, that is a more difficult document than any forecast.
