AI and Biosecurity

This working paper analyses the development of Large Language Models LLMs and AI-enabled biological tools in the biosecurity area and presents the risks and benefits with practical examples. Artificial Intelligence (AI) is commonly understood as the ability of machines to perform tasks that normally require human intelligence. In the biosecurity area, the AI-enabled biological tools have a similar rate of progress and model size expansion as the LLMs resulting in reduced requirements for time, money, expertise, and equipment. Many AI-enabled biological tools are open-source and can be downloaded and modified by users; applications include protein design models, genomic language models, autonomous laboratory robotic and AI agents. The key problem is the dual-use potential of AI-enabled biological tools which can be both beneficial and harmful. For example, tools that detect new viruses could increase awareness, but also identify new candidate viruses for bioweapons. The same is applicable for mutation prediction tools which are helpful for pandemic preparedness and vaccine development, but could also give hints how to create viruses that can evade from countermeasures and the immune system. The progress of AI is reducing the current hurdles of tacit research knowledge and physical access to harmful material. LLMs such as GPT-4 and 4o can provide step-by-step plans for creating chemical and biological weapons. The o1 model responded correctly to questions about the release of bioweapons of 80% and produced plans superior to PhD-level experts. Advanced models like OpenAI’s o3 and Google’s Gemini 2.5 Pro have outperformed 94% of PhD-level virologists in the Virology Capabilities Test. In a course at the Massachusetts Institute of Technology, chatbots suggested four potential pandemic pathogens, explained how they can be generated from synthetic DNA, provided names of DNA synthesis companies unlikely to screen orders, identified detailed protocols, and recommended to engage reverse genetics facilities. MegaSyn is an AI-based drug discovery program which designed highly toxic nerve agents after inversion of the program logic. The advanced protein structure analyzer AlphaFold 3 can predict the structure of complexes while Evo2 is a step towards designing completely new genomes. Meanwhile, multi-agentic research systems like the Virtual Lab are available. On the other hand, AI is already supporting pathogen surveillance systems, the development of medical countermeasures and of pandemic preparedness and response, because AI systems can recognize patterns for early detection. Advanced biodefense tools are FunGCAT for DNA sequence identification and HoloZScan for pathogen detection and identification. OpenAI plans a biodefense summit in July 2025. Already active countermeasures include monitoring of the AI models by always-on detection systems. combining automated systems with human reviewers with investigations in case of suspicious activities. In summary, the paper has shown the rapid progress of LLMs and AI-enabled biological tools and the growing dual-use potential of these technologies. The impact on biosecurity is underestimated by policymakers which led to significant regulation gaps which need to be addressed urgently.