AI Tools Used to Plan Dangerous Virus Experiments

Anthropic has released a threat report detailing how researchers in other countries used its AI models to plan experiments with high-risk viruses. The company identified five specific instances where AI was used to study how viruses evade immune systems and increase transmissibility, raising serious concerns about the potential for bioweapon development.

According to the report, these researchers did not just ask general questions. They used the technology to design novel toxins and modify viruses to make them more dangerous. This moves the debate on AI safety from hypothetical risks to documented, real-world attempts to bypass digital safeguards and accelerate sensitive biological work.

Researchers Bypass Digital Safety Barriers

One notable case involved a military research institute attempting to study chikungunya. The goal was to identify mutations that would make the virus spread more easily and evade the body’s defenses. While such research can lead to better vaccines, it also creates a pathogen that is harder to distinguish from a natural outbreak.

To access the AI, the researchers used a platform that routed their traffic through American infrastructure, effectively hiding their location. When the AI refused to answer sensitive prompts, the platform switched to a different model with looser restrictions. Even after Anthropic banned the accounts, the researchers regained access within days using new identities, showing how difficult it is to permanently block malicious use.

Bird Flu Studies Raise Pandemic Fears

In another incident, a researcher spent weeks using AI to plan experiments on highly pathogenic bird flu. The focus was on making the virus adapt to mammals and spread through the air. A version of bird flu that transmits easily between humans would pose a significant pandemic threat.

The report suggests the researcher had access to virus samples and animal testing facilities. Anthropic limited the AI’s role in this case to study design and data analysis, but the company noted that the work represented an active program to develop dangerous biological knowledge. This highlights a key trade-off: AI can help scientists, but it can also give them tools to create threats.

New Reality for AI Safety Debates

This report adds weight to ongoing discussions about catastrophic AI risks. While some voices call for pausing development, Anthropic’s data shows that the danger is not just about future superintelligence. It is about present-day actors using current technology to circumvent geographic and technical restrictions.

As reported by GN technics/ai (en-US), the situation underscores a clear challenge for AI developers. They must find ways to prevent dangerous applications without slowing down beneficial scientific research. The incident with the chikungunya virus shows that simple bans are not enough when determined actors can find workarounds.