by Charlie Harrison
Synthetic biology, the engineering of new organisms or bioactive compounds, is one area where AI could have a truly transformative impact. In this sixth and final episode of the AIBIO podcast series, the hosts discuss the threat of AI producing a weapon of mass destruction – and the guardrails that we rely on to prevent this from happening. Listen to the episode in full here or wherever you get your podcasts.
A key idea dilemma is that technological breakthroughs often have multiple uses. The previous episode discussed AI-powered ecological monitoring, but the same technology can be used for mass surveillance. GPS was originally a military development before it made its way into commercially available satnavs and phones; and the techniques behind self-driving cars also power autonomous weapons. Nowadays, the same computational methods used to design novel drug candidates can be easily misused to identify toxic compounds.
Not long ago, designing and editing genome sequences was extremely challenging time consuming, and expensive. Advances in bioinformatics, along with robotic laboratories, has dramatically reduced the time it takes to design, build, and test synthetic biological compounds or organisms. The leading edge in biotechnology is moving extremely fast – much faster than the regulations designed to govern it – and the natural barriers to innovation, whether benign or malicious, are being challenged by AI.
To create a new synthetic biology product – a drug, a treatment, a biomaterial, or a toxin – you first need an idea. This requires deep domain knowledge. Modern frontier-model chatbots, from the likes of OpenAI and Anthropic, can provide this in abundance. The leading models now have built-in filters that flag potentially dangerous discussions, but their reliability and their resistance to jailbreaking remains a risk. Open-weight models, while less capable than the frontier, can be deployed without guardrails by anyone with sufficient computational hardware.
Next, the idea needs to be refined and assessed in terms of its biological function. We now have extremely powerful open-source AI tools, like AlphaFold, that can simulate the structure and function of proteins. Many of these tools have been assessed to have a high potential for malicious use.
The physical creation of biomaterial is another barrier. The cost of DNA synthesis has come down by three orders of magnitude in the last two decades, and there are now online mail-order services that will ship custom DNA sequences ready to implant in an appropriate organism. These companies screen the sequences that are submitted to them to avoid producing materials from, for example, smallpox or anthrax. In 2023, a research team showed that they could use open-source AI tools to redesign malicious proteins to produce functionally equivalent versions that beat the screens. More recent tests across DNA synthesis providers in July 2026 found that screening successfully detected 96.75% of malicious sequences – meaning that more than 3% still get through. This screening, along with identity checks, is currently completely voluntary.
Finally, there is the transition to real-world applications. Operating a laboratory is still hard. However, AI combined with robotics is lowering the barrier to entry and making it possible for relatively small groups to participate in this last step.
How can society protect itself against this threat? It might be that our only chance is “defensive acceleration”: invest in the technologies that protect us from bioweapons. Doubtless these will also rely heavily on AI.