Anthropic Says It Blocked AI Help for Biological Weapons

Anthropic says it blocked attempts by harmful actors to use its AI models for work that could have supported biological weapons.

The disclosure came in its September 2026 threat-intelligence report, a public record of misuses the company says it stopped. The report covers December 2025 to August 2026 and lists seven types of harm, from cyber operations to biological misuse. Reporting on the findings on Sept. 11 said the blocked biological activity could have supported weapons development, a claim that drew international attention PBS.

After the report, experts called for tighter limits on access to powerful AI models, citing rising misuse cases including biological research risks Al Jazeera. They did not call for a small fix. They called for controls on what models can do, how they are released, and who can use them, much like safety rules around who can enter a lab and what tools they can use.

This warning follows earlier disclosures. Scientists shared transcripts with The New York Times in which AI chatbots explained how to put together deadly pathogens, or disease-causing germs, and release them in public places The New York Times. In a separate case, biology and terrorism experts who reviewed conversations with ChatGPT described some answers as deadly accurate The Wall Street Journal.

A new study added to concerns that AI could help create a new generation of biological weapons The New York Times. That finding fits into a debate reported in Nature about whether to limit biological AI, as researchers warn AI can design viruses, toxins and other bioweapons. Researchers have also found that AI-designed toxins can pass through safety screening used by companies that sell genes, the ordered DNA pieces labs use for research.

Governments have followed this risk for years. In July 2023, Democratic and Republican U.S. senators raised bipartisan alarm about AI, with a focus on the risk of a biological attack. The U.S. intelligence community later gave its formal view. The ODNI 2026 Annual Threat Assessment states that AI has the potential to aid in weapons and systems design.

In June 2026, Anthropic released a proposed policy framework for governments to address four risk categories for AI models, including biological weapons. In an hourlong interview, Bill Gates said separately that the tech industry was downplaying AI risks, which he said include mass unemployment and bioterrorism The New York Times.

The broader context here is a shift from possible misuse to cases that were stopped. Early warnings focused on what models might do in theory. These reports describe what actors tried to do, what companies caught, and where gene screening failed. That changes the policy question from forecasting ability to handling incidents, sharing information, and setting lines for action.

Looking at what this means for diplomacy, disagreements are easy to map. States will differ on verification, on business rules for models and gene synthesis, and on how much threat data private firms should share. Companies hold usage records. Governments hold regulatory power. Researchers hold knowledge of pathogens and toxins. No group can solve this alone. The test will be whether they can build a shared base of evidence before the next report describes a more advanced attempt.