The public debate about AI and weapons risk often begins with chatbots. But the sharper concern, according to researchers in bioterrorism and health intelligence, sits elsewhere: specialized AI systems built for the life sciences.
Those tools can support disease research and drug discovery. They can also create harder questions for biosecurity because useful scientific systems may be usable in more than one way.
Chatbots are not the main risk
Large language models, or LLMs, power chatbots such as ChatGPT, Claude and Gemini. They can summarize information, answer questions and help users move through existing material more efficiently.
That matters, but it is different from inventing a new biological threat. The source article argues that today’s chatbot systems draw on already-existing data and are unlikely to produce genuinely new capabilities by themselves.
The example of OpenAI’s o1 model shows why risk language can be easy to misread. In September, OpenAI released o1, nicknamed “Strawberry”. The developers described the model as having a “medium” level risk of helping someone create a biological weapon.
On its face, that sounds severe. But the article says a closer look at the o1 system card points to less dramatic issues. One example is that the model could help an untrained person search a public database of genetic information about viruses more quickly.
That kind of assistance may be useful, but the researchers argue it is unlikely to change the biosecurity picture in a material way. It may help someone organize ideas or find a starting point, yet that is not the same as enabling the creation of a weapon.
The harder problem is dual-use AI
The more serious challenge comes from AI systems designed for scientific work. The article points to life sciences tools, including the AlphaFold series, as systems that can help researchers fighting diseases and looking for new therapeutic drugs.
That benefit is central to the problem. A system that is powerful enough to help science may also be powerful enough to be misused. The article describes this as “dual-use research of concern”.
Dual-use risk is not new. Biosecurity and nuclear non-proliferation specialists have long dealt with the fact that tools developed for legitimate research can also support harmful goals. Chemistry and synthetic biology already contain many techniques that could be directed toward malicious ends.
AI adds a new layer because it may make scientific exploration faster, more accessible or more capable. The article does not claim that this danger is settled. Instead, it stresses that the exact implications for “chem bio” weapons remain uncertain while regulation is failing to keep up with technical change.
Prions, toxins and pandemic potential
Protein science shows how the issue can become concrete. The article notes concern for more than a decade that computational platforms could help with the synthesis of prions, which are potentially deadly misfolded proteins, or with the construction of novel toxin weapons.
New AI tools such as AlphaFold may bring that scenario closer to reality. Even so, the article draws an important distinction: prions and toxins may be deadly to relatively small groups people, but neither can cause a pandemic that could wreak true havoc.
For bioterrorism researchers, the highest concern is with agents that have pandemic potential. The article names Yersinia pestis, the bacterium that causes plague, and variola virus, which causes smallpox, as historical focuses of bioterrorism planning.
The central unresolved question is whether new AI systems make a practical difference for an untrained person or group trying to obtain such pathogens or create something from scratch. The article’s answer is blunt: right now, we simply do not know.
Assessment rules are still immature
One major gap is that there is no consistent, widely followed method for assessing the biological weapons risk of an AI model. That makes public claims hard to compare and easy to exaggerate.
The most developed planning described in the source article comes from the outgoing Biden administration in the United States. An executive order on AI development issued in October 2023 directed several US agencies to establish standards for assessing how new AI systems may affect the proliferation of chemical, biological, radiological or nuclear weapons.
Experts often group those categories as “CBRN”. The article uses “CBRN+AI” to describe the still-uncertain overlap between these weapons risks and artificial intelligence.
The executive order also created new processes for regulating the hardware and software involved in gene synthesis. That matters because gene synthesis can turn digital biological designs into physical biological material.
The US Department of Energy is also soon due to release guidance on managing biological risks that might be generated by new AI systems. According to the article, that guidance could offer a path for understanding how AI may shape biosecurity in the years ahead.
Politics may reshape the response
The developing regulatory approach is already facing political pressure. The incoming Trump administration in the US has promised to repeal Biden’s executive order on AI, saying it is based on “radical leftist ideas”.
The source article argues that this framing is disconnected from the biosecurity question. Whatever the executive order’s flaws, the article presents it as the best existing blueprint for understanding how AI could affect chemical and biological proliferation.
The practical issue is not whether every fear about chatbots is justified. Many are likely overstated. The real test is whether governments can assess and manage specialized life sciences AI before scientific capability moves further ahead of oversight.
That makes the future of AI biosecurity less about chatbot panic and more about disciplined evaluation. The risks are uncertain, but the need for serious rules is already visible.