On September 10, Anthropic announced that it disrupted five cases involving AI and biological weapons, after users reached Claude through accounts, private networks, and third-party services to seek help with research that enables dangerous biological capabilities in practice worldwide.
The cases do not prove Claude produced a working weapon.
Safeguards and Dual-Use Biology
Anthropic’s report describes five cases in which its models could support malicious biological research. The work involved chikungunya, bird flu, smallpox, and animal toxins.
One military-linked researcher used Claude for a chikungunya proposal. Anthropic said the work could support vaccine development, yet, some requests might also help create more dangerous virus variants.
That overlap centers AI biological weapons.
“The same information that can be used to develop a biological weapon could also be used to develop, for example, a vaccine or a cure for a disease,” claimed Anthropic.
Users also tried to bypass safeguards with fake accounts, private networks, and third-party sellers, making detecting malicious AI agent activity a problem involving model filters, identity checks, access controls, and monitoring.
Anthropic said access to some biology-capable models is restricted to examined organizations, which adds another layer around biological weapon research but alo reveals that safety depends on who can reach frontier systems.
Claude was also used for missiles, drones, bombs, firearms, and targeting systems.
In one case, the report described a request originating from Yemen, involving the development of guided rockets using mobile technology, pushing AI and biological weapons into a broader security problem.
Warnings from Inside the Industry
Anthropic’s controls look more urgent against malicious AI agents that were also used in cyber espionage, propaganda, surveillance, scams, and fraud.
Google reported an attempt to use Gemini as a guide to synthesizing weaponized biological agents, suggesting that malicious biological research is becoming a recurring test for frontier AI companies.
Meanwhile, the real challenge comes from Anthropic.
Evan Hubinger, a safety researcher, said there is a greater than 10% chance advanced AI could kill all humans within the next decade, though current risk remains low.
“I believe Anthropic is trying its best, but we do not yet have a plan to solve alignment for superintelligence and are not clearly on track to,” said Hubinger.
Hubinger’s warning places AI and biological weapons inside a broader fear that capability may be advancing faster than control.
“Neither company is acting responsibly,” wrote former Anthropic and OpenAI researcher, Jacob Coxon.
Coxon warned future systems that may hold the capability to hack systems, transform fields quickly, and gain power and resources. Such claims deepen concern about malicious AI agents.
Those warnings are now reaching policymakers. US Senator Bernie Sanders has called for a pause on advanced AI development and a ban on artificial superintelligence, reflecting growing concern that governments need to act before more powerful systems emerge.
President Donald Trump has rejected that approach, arguing that slowing development could weaken the US in the global AI race. Speaking Thursday, he said he was concerned if they don’t win AI, they are going to be put in a very bad position.
With minimal human supervision, dangerous behavior may become harder to detect before damage begins. The phrase AI bots told scientists how to make biological weapons goes further than Anthropic’s evidence supports.
The company declared its cases show potentially dangerous assistance, not proof that users-built weapons. Additionally, Anthropic stressed that limit.
“Evaluation is useful because it provides evidence of the capabilities of a system. However, evaluation cannot concretely show that these capabilities will be used to develop biological weapons in the real world,” said Anthropic.
Still, the gap between capability and intent is where AI and biological weapons become difficult to govern, since dual-use science can blur harmless and harmful requests. That is also the problem with AI agents misuse at scale.
If users hide behind new accounts, guardrails may act more like checkpoints than permanent barriers.
Anthropic’s report show safeguards can catch real cases. But its employees’ warnings suggest AI biological weapons cannot be treated as a solved safety problem.
The question is whether biological weapon research can remain containable as models become more autonomous, and Anthropic’s interventions matter, while internal warnings suggest today’s controls may only buy time.
That leaves AI and biological weapons as a test of whether AI companies can prove their safeguards are stronger than the systems they build.
Inside Telecom provides you with an extensive list of content covering all aspects of the tech industry. Keep an eye on our Impact section to stay informed and up-to-date with our daily articles.
