Anthropic says it has thwarted several potential attempts to use artificial intelligence in the development of biological weapons during 2026, raising fresh concerns about the misuse of advanced AI systems.
The company said some users had employed its Claude models to carry out research that could have supported the creation of biological weapons. Anthropic said it blocked the activity and took action against the accounts involved.
It said it was not always possible to establish whether the work had a legitimate scientific purpose or a criminal intent, because the same biological research can be relevant to medical advances as well as the development of dangerous pathogens.
Anthropic has introduced tighter safeguards around biological queries as the capabilities of its models have increased. The company’s own testing found that access to an unrestricted version of Claude improved participants’ performance on tasks related to bioweapons planning and acquisition compared with people using basic internet resources alone.
The firm has said its models are subject to real-time monitoring and access controls designed to prevent assistance with chemical and biological weapons. It has also warned that the technology’s ability to connect information across different areas of biology could save significant time for people with malicious intentions.
The disclosure comes as former and current Anthropic researchers have intensified warnings about the wider risks posed by increasingly powerful AI. One departing researcher said the companies developing the technology were moving too quickly and warned that some people within the industry believed advanced systems could ultimately threaten humanity.
