Anthropic has blocked an attempt to use its artificial intelligence models for biological research that could potentially have increased the harmful properties of chikungunya virus, highlighting the growing challenge of controlling AI systems capable of assisting with advanced scientific work.
The US-based AI company said researchers identified the case as part of a broader investigation into misuse of its models between December 2025 and August 2026.
The incidents ranged from politically motivated activity and state-linked propaganda operations to attempts by spyware companies to exploit AI capabilities.
But the biological research case illustrates a more difficult problem for AI developers: the same systems that can accelerate legitimate scientific research may also be useful in activities that create public-health and security risks.
Anthropic says AI safety threshold is changing
According to Anthropic, a user attempted to obtain assistance with a scientific funding application involving research intended to create a new or enhanced biological property of chikungunya virus.
The company determined that the request could potentially have been used to make the pathogen more dangerous and blocked it.
Chikungunya is a mosquito-borne viral disease that can cause fever and severe joint pain. The concern raised by Anthropic was not simply that its AI could provide information about an existing pathogen, but that increasingly capable models could assist with more sophisticated biological research.
That distinction is becoming important as AI systems move beyond general-purpose question answering and into complex scientific tasks.
Anthropic said its older models were previously considered below the threshold at which they could meaningfully assist users with dangerous biological research.
That assessment is becoming less certain as newer models become capable of supporting researchers across increasingly complicated scientific workflows.
Why advanced AI changes the risk
AI companies have historically focused many biological safety controls on preventing models from providing information that could help inexperienced users reproduce known biological threats.
Anthropic says that approach may no longer be sufficient.
The company has introduced stronger restrictions covering a broader range of biological research involving dual-use applications — work that can have legitimate scientific purposes but could also be misused.
The change reflects a wider concern across the technology industry and among policymakers: AI models do not need to independently conduct an experiment to create a security risk.
Helping a user analyse research, identify approaches, organise scientific work or overcome technical obstacles can itself become consequential when the underlying research involves dangerous biological systems.
Anthropic reports wider misuse of its models
The biological research attempt was one part of a broader set of cases investigated by Anthropic.
The company said it had identified misuse involving politically motivated individuals, groups seeking to spread propaganda and commercial spyware companies.
It also reported an industrial-scale attempt to covertly extract the capabilities of one of its models and reproduce them in another system without authorisation.
Anthropic said the biological research incident and the other cases covered in its report did not involve its newer Claude Fable or Mythos-class models, apart from the model-extraction campaign.
The company said its stronger safeguards are intended to account for the capabilities of newer generations of AI rather than relying solely on restrictions designed for earlier systems.
The regulation problem is becoming harder
Anthropic’s disclosure comes as governments face growing pressure to determine how advanced AI should be regulated.
The central difficulty is that the risks are evolving alongside the technology.
A model that poses limited biological risks today could become substantially more capable after relatively small improvements in reasoning, scientific knowledge or its ability to use external tools.
That makes static safety rules harder to maintain.
AI companies have therefore increasingly become responsible for deciding which scientific requests their systems should answer, which they should restrict and which should trigger further scrutiny.
Critics argue that such decisions should not be left entirely to technology companies.
The debate has become particularly urgent as AI developers pursue increasingly autonomous and capable systems.
One Anthropic researcher, Jacob Coxon, announced his resignation earlier this week and criticised the direction of the industry, arguing that companies including Anthropic and OpenAI are moving rapidly towards self-improving AI systems while taking potentially significant risks.
His comments underscore a broader disagreement within the AI industry over how quickly frontier models should be developed and what level of risk society should accept.
For Anthropic, the chikungunya case offers a concrete example of why that debate is no longer theoretical.
The question is no longer only whether AI can help scientists discover new things. It is also whether increasingly capable models can distinguish between scientific progress and research that could create new threats — and whether the safeguards around those models can keep pace.
