Anthropic says it has shut down accounts after finding people using its Claude AI models for activities that raised concerns about biological weapons, cyberattacks and other forms of misuse.
The company released the findings Thursday in a threat intelligence report covering activity from December 2025 through August 2026. It described users ranging from scientists and suspected government-backed groups to criminals, propaganda organizations and spyware companies. Anthropic said it responded by blocking accounts and adjusting its safety systems.
Five cases involved biological research that the company believed could potentially be useful in creating biological weapons. Anthropic said determining the users’ intentions was difficult. Some scientific information has more than one possible use. Work that contributes to medicine or vaccine research, for instance, could also provide knowledge useful for making a dangerous pathogen.
One case involved research into chikungunya. A scientist turned to Claude for assistance connected with a funding proposal for work on the virus. The planned experiments involved altering it and observing what happened as it moved through animals. Anthropic said the situation raised greater concern after it found that the research was connected to a military research facility.
The company did not reveal who the scientist was or identify the facility or country. According to the report, people involved in some of the biological cases also tried to disguise what they were researching so Claude’s restrictions would not stop them. Anthropic eventually closed the accounts.
This is becoming a harder problem as AI systems improve. Anthropic said testing of older models last year suggested they could not offer much meaningful help with dangerous biological work. Newer models are able to perform more complicated scientific tasks, leading the company to expand its controls around biological questions that could have both legitimate and harmful uses.
The report went beyond biology. Anthropic found six cases where Claude was used for computer programs connected to conventional weapon systems. The work covered several kinds of weapons and related control or targeting technology. Those cases were traced to users in China, Russia and Yemen.
Cyber operations appeared in the report as well. Activity matching the methods of the Russia-based hacking group Midnight Blizzard allegedly involved using AI to help malware respond when security systems detected it. The software could then be changed in an attempt to get past those defenses.
Anthropic also reported suspected surveillance and propaganda operations. Chinese and Iranian-linked actors were associated with some activity targeting dissidents or communities living outside their home countries. Russian state media allegedly used Claude to produce material that looked like independent news coverage, including false information involving an election in Moldova.
The company is not alone in reporting attempts to use AI for potentially dangerous biological work. Google said this week that a user tried to get its Gemini system to provide detailed instructions related to weaponized biological agents.
Anthropic said information gathered from these incidents is now being used to strengthen its methods for finding and stopping misuse. It has also provided information to government authorities and other companies in cases where it considered sharing appropriate.
This image is the property of The New Dispatch LLC and is not licenseable for external use without explicit written permission.
