Anthropic says it blocked misuse of its AI that could have aided biological weapons work

Anthropic said it blocked attempts to use its artificial intelligence systems for research that could have supported biological weapons development, part of a broader set of misuse cases the company said it uncovered between December 2025 and August 2026. The findings matter because they highlight how frontier AI tools can be steered toward dangerous dual-use biology tasks even when safeguards are in place.

In one case, Anthropic said its systems stopped a request for Claude’s help with a grant application tied to gain-of-function research on chikungunya virus, a mosquito-borne illness. The company said the work involved attempts to make the virus more transmissible and better able to evade immune responses, while noting that such research can also support legitimate vaccine and treatment development.

Anthropic described the incident as one example from a larger set of nine misuse cases it said it identified across regions including Russia, Iran, Turkey, the Persian Gulf, South Asia, Africa and Europe. The company said the cases involved actors ranging from spyware vendors and politically motivated individuals to state-backed groups spreading propaganda.

A broader pattern of misuse

The company said the report is its third since March 2025 documenting abusive uses of its technology. Anthropic said it published the findings because it believes AI developers have a responsibility to disclose malicious use of their systems and to help governments and competitors detect similar activity.

Anthropic has made biology safety a central part of its public safety messaging. In a recent company post, it said its newer Fable 5 model launches with most biology queries blocked because of the risk that advanced capabilities could be misused to develop a biological weapon. The company said it relies on automated safety classifiers and other controls to distinguish benign from harmful requests, while acknowledging that the systems can still generate false positives and are subject to jailbreak attempts.

The company’s transparency materials say it evaluates models for biology, cybersecurity and autonomous capabilities, including whether a model can help create, obtain or deploy biological weapons. Anthropic’s responsible scaling policy also calls for safety evaluations before frontier models are released in areas linked to chemical, biological, radiological and nuclear risk.

Anthropic said the misuse cases it identified were not typical everyday abuse, but among the most notable and novel threats it has seen so far. The disclosure comes as policymakers and researchers continue to debate how quickly advanced AI systems should be deployed, and what guardrails are needed as the technology becomes more capable.