Anthropic reveals five biological weapon risk cases
Anthropic says it has caught scientists using its AI models for research that could help develop biological weapons, including work aimed at making dangerous pathogens more transmissible and harmful.
Five cases of biological researchers using Claude raised serious security concerns, according to a company threat intelligence report published on Thursday.
One case, to be carried out at a military research institute, involved gain-of-function research aimed at making the chikungunya virus more transmissible and better able to evade the immune system.
Another scientist used Claude while researching highly dangerous bird flu and ways the virus could adapt to mammals. Other cases involved work on viruses related to smallpox and the redesign of dangerous toxins.
Anthropic said some of the researchers went to considerable lengths to get around its restrictions, including hiding their identities, routing traffic through other countries and using services designed to bypass safety controls. The accounts involved have been banned while the services used to get around the safeguards have been shut down.
The company stopped short of saying the scientists were definitely trying to build biological weapons. However, it warned that sophisticated actors may deliberately hide malicious aims behind legitimate-looking research.
Anthropic acknowledged that its AI is becoming more capable, saying older models were too weak to meaningfully assist sophisticated users with dangerous biological research, but that it can no longer make the same assurance about today’s systems.
The findings come just weeks after US scientists used AI to design complete viral genomes and successfully created 16 functioning viruses that did not previously exist in nature. Although those viruses infected bacteria rather than humans, researchers have warned that the experiment demonstrates how quickly AI’s biological capabilities are advancing.
Anthropic’s report also detailed other alleged misuse of Claude, including cyberattacks, surveillance, propaganda, fraud and assistance with conventional weapons such as missiles, armed drones and bombs.
Just days before the threat assessment was published, Anthropic researcher Evan Hubinger publicly stated that he believes there is a greater than 10% chance AI could “kill all humans” within the next decade, adding that the company still has no clear plan for safely controlling superintelligence.
His admission came after another Anthropic researcher, Jacob Coxon, resigned, accusing Anthropic and OpenAI of ignoring the “civilizational stakes” and “racing straight to self-improving superintelligence and gambling with our lives.” He separately warned that sufficiently advanced AI could hack critical infrastructure or help create “extinction-level bioweapons.”
