Anthropic says it blocked five cases of Claude use that could support biological weapons

Anthropic says it identified and disrupted attempts to use its Claude models in ways that could support the development of biological weapons. The disclosure sits inside a wider threat intelligence report covering malicious activity the company blocked between December 2025 and August 2026, according to the BBC.

What the report found

Anthropic documented misuse of its Claude Haiku, Sonnet and Opus models across an eight-month window, with the biological cases as the headline: five case studies of actors using the models in ways that could support biological weapons development. The California-based company described biological misuse as among the most serious risks of frontier AI, warning that without the correct safeguards such capabilities "could have catastrophic consequences".

It also confronted the dual-use problem directly, noting that the same information that feeds a weapon could equally feed a vaccine or a cure. Jacob Klein, Anthropic's head of threat intelligence, told the New York Times it was "an incredibly nuanced situation". As he put it: "You are not seeing someone in a comic book kind of way say, 'Hey, I want to build a biological weapon to kill everybody.'"

Beyond biology, the report listed six cases in which Claude was used to develop software for conventional weapons — firearms, missiles, armed drones, bombs and other munitions — along with the targeting and control systems that operate them.

The actors involved

The report names suspected state-sponsored groups, criminal hackers, spyware vendors, state propaganda institutions and politically motivated individuals. Claude was used by actors linked to a Russia-based cyber espionage campaign and by an Iranian propaganda institution, the BBC reports. Other activity flagged during the period ranged from fake dating apps and hotel Wi-Fi scams to surveillance tooling built to identify dissidents.

A group whose behaviour is consistent with the Russia-based Midnight Blizzard allegedly used AI to build a system that spotted when security defences flagged its malware and kept rewriting the code until it slipped past detection. The hacking group ShinyHunters and China-based labs were also named, and Anthropic separately accused Chinese AI firms of trying to replicate Claude's capabilities. None of the documented misuse involved Claude Fable or the more powerful Mythos-class models, the report said, apart from a single instance of distillation — a technique in which smaller AI models are trained using larger, more expensive ones.

A week of warnings

Anthropic's disclosure landed amid a run of stark statements from industry and politics. Google disclosed on Tuesday that a person had attempted to use its Gemini model to obtain a "complete, step-by-step technical guide for synthesizing weaponised biological agents". OpenAI chief scientist Jakub Pachocki, writing on 6 September, called for "voluntary slowdowns" until safeguards are in place, arguing that nobody is ready for what a continued rapid rise in machine intelligence would bring.

In the US, Senator Bernie Sanders has introduced legislation to ban artificial superintelligence and temporarily pause advanced AI development, telling BBC Newsnight that when scientists warn of a chance of cataclysmic impact, "you've got to be a moron not to say, slow it down". President Donald Trump has rejected such fears, saying he was concerned that "if we don't win AI, we're going to be put in a very bad position". In the UK, the warnings prompted an open letter to Prime Minister Andy Burnham calling for a multinational treaty on the safe development of AI.

Why it matters

Anthropic says it has folded its findings into its own prevention, detection and disruption processes and shared intelligence with authorities and industry partners where appropriate. Even so, the report stands as one of the most concrete public accounts yet of frontier models being probed for weapons-relevant work by named criminal and state-linked actors, rather than as a hypothetical risk exercise. It also frames the core dilemma of AI safety in practical terms: the same capabilities that accelerate medicine can shorten the path to a weapon, and telling the two apart can hinge on intent that is never stated outright. How hard labs police that line — and whether governments respond with treaties or development pauses — is likely to shape the next phase of the AI policy debate.