Anthropic caught researchers using Claude to potentially develop biological weapons
An Anthropic report reveals extensive misuse of Claude.
Biological warfare powered by Claude. Image by Cybernews
Follow us
- Anthropic says it found five cases where users asked Claude to support biological weapons work.
- One case involved research to make bird flu spread more easily between mammals.
- Some users tried to bypass Claude’s safety rules by routing requests through unsupported regions.
- Anthropic also flagged venom research that could help make paralytic drugs.
Key Takeaways by nexos.ai, reviewed by Cybernews staff.
Anthropic has caught researchers and scientists using Claude to work on substances that could aid in biological warfare.
Anthropic has just released a report detailing numerous cases of users exploiting its AI model Claude.
While bad actors continue to use AI for influence operations and state-sponsored activities in certain countries, other, more obscure use cases have been brought to light.
“Biological misuse is one of the most serious risks of frontier AI models,” Anthropic said, echoing widespread fears that AI models may one day manipulate pathogens to make them more dangerous.
Models like Claude Opus 4 and Claude Sonnet 4.5 (both released in 2025) were “well below the threshold” when it comes to aiding the manipulation of toxic substances, but more powerful models like Claude Fable 5 are a different story.
Anthropic identified 5 case studies in which actors used its models to support the development of biological weapons.
Human bird flu variant research flagged by Anthropic
The first example involved a reseller platform that circumvented geographic blocks to help virologists working on the virus “Chikungunya,” which is spread by a specific type of mosquito.
Virologists were apparently ordered by the state to genetically alter or enhance the virus, with a focus on transmissibility and immune evasion.
The researchers asked Claude for help, but the AI model refused. So the actors funneled traffic through different regions to evade detection and targeted models with fewer safeguards.
Anthropic blocked the users associated with the requests and banned future requests from that account.
Other case studies involved highly contagious viruses, such as orthopoxviruses, which include infections like smallpox and monkeypox.
One notable case involved the highly pathogenic avian influenza, otherwise known as bird flu.
The research focused on adapting the flu to be more transmissible between mammals. Bird flu doesn’t spread effectively from person to person.
So, the development of a variant which could be transmitted between humans is “of very high concern” as if it were transferred to humans, it would be almost 100% fatal.
Like the first case, the actor accessed Claude from a region of the world where Anthropic’s models aren’t supported.
Potential locations include mainland China, Hong Kong, the administrative Chinese region of Macau, Russia, and other areas under Western sanctions.
Stay updated with our latest stories and follow us on social media
Be the first to discover new stories, ideas, and updates from our team.
Claude users attempt to exploit venoms
But it's not only viruses that are being exploited by bad actors – venoms and toxins are also on the table.
While some of the compounds researched by the actors could have dual uses, it appeared that the person wanting more information didn’t have good intentions.
In its report, Anthropic notes that research can be used for dual purposes.
Anthropic caught a researcher using Claude to develop an atlas of venom toxin peptides from different lineages of venomous animals.
The actor specified that they were researching this for a therapeutic goal of developing painkillers and antidepressants.
However, Anthropic saw an issue with this, as the atlas could also be used to develop paralytic drugs, which are powerful drugs used to temporarily paralyze muscle movement.
These drugs could then be used to chemically submit a person or even kill them.
