Chemical AI Has a Weapons Problem Nobody Planned For

The Chemical Weapons Convention Was Built for a World Where Synthesis Knowledge Was Rare. AI Changed That.

On May 20, MIT published a profile of Connor Coley, a 31-year-old MIT associate professor who builds AI models for drug discovery. The Chemical Weapons Convention has existed for 28 years. The treaty was built on the assumption that synthesizing chemical weapons requires rare expertise and access to controlled precursors — the kind of expertise that takes years to acquire and isn't available from a laptop. Coley's work suggests that assumption is now obsolete. The treaty has no plan for what comes next.

Coley, the Class of 1957 Career Development Associate Professor of Chemical Engineering and EECS at MIT, has spent years building models that predict chemical reactivity, evaluate drug candidates across 3D molecular shape and electrostatic interactions, and navigate synthesis planning with a 71 percent success rate across 35 validated novel compounds. His lab's FlowER model, published in Nature in August 2025, conserves mass in ways that make it particularly useful for predicting reaction pathways. His ShEPhERD model is now in use at pharmaceutical companies hunting for viable drug candidates. These are legitimate, valuable tools. They are also, by definition, dual-use.

The technical shift that makes this a non-proliferation problem is specific. Early molecular AI worked with 1D SMILES strings or 2D message-passing networks, producing molecules that looked good on screen but could not necessarily be synthesized or were practically infeasible. The field has moved to synthesis-aware-by-design, where models generate molecules that are not just chemically valid but actually producible in a lab. The SYNC 3D benchmark, discussed at ICLR 2026, is replacing older metrics like SA-Score, which captured only 2D synthesizability. This is progress in drug discovery. It is also a capability that was previously inaccessible.

What makes this structurally difficult for the Chemical Weapons Convention is that the treaty operates through declaration and inspection. Member states declare their chemical activities and permit inspections to verify compliance. This model assumes that a weapons program leaves traces: large facilities, specialized equipment, precursor chemicals (the controlled starting ingredients needed to run a synthesis reaction). A novel compound can be designed without ever touching a regulated precursor. A synthesis route can be computed for compounds that have never been made, using standard laboratory reagents. The detection problem changes fundamentally when the threat is not a barrel of nerve agent but a data file and a reaction flask.

The Organisation for the Prohibition of Chemical Weapons flagged exactly this in a landmark March 2026 report from its Scientific Advisory Board. The working group found that AI lowers barriers across multiple stages of chemical weapons development: identifying candidate compounds, planning synthesis routes, even designing molecules that evade existing detection methods. The dual-use problem is not incidental to these systems — it is a consequence of how they work. The report called for continued monitoring and technical engagement but did not propose specific regulatory mechanisms. The treaty was written for a world where chemical weapons programs required industrial facilities and traceable precursors. It has no language for software that generalizes across chemical space without being told what to do with that generality.

The International AI Safety Report, published in February 2026 by more than 100 experts citing over 1,400 sources, reached a similar conclusion. Chemical weapons AI synthesis was flagged as an increasing and emerging concern, not a speculative one. Earlier that same month, a Nature paper described the MOSAIC system: 2,498 expert models navigating synthesis planning with a 71 percent success rate, validated experimentally across 35 novel compounds. The system runs on Llama 3.1-8B, open-source and commodity hardware-accessible. The authors released their weights publicly.

The numbers involved in chemical space are difficult to reason about. Potential small-molecule drug-like compounds are estimated between 10^20 and 10^60, a range that reflects genuine uncertainty about the boundaries of synthetically accessible chemistry. The MOSAIC authors are not the only group working on this problem, and the field is moving faster than any single institution can track.

Coley has argued publicly that the dual-use problem is real and that the technical community has a responsibility to engage with policy. There is no natural boundary where a drug discovery model becomes a weapons design tool. The same architecture, the same training data, the same molecular representations apply across both.

The Chemical Weapons Convention rests on the proposition that some forms of knowledge are too dangerous to leave unrestricted. That proposition is now in tension with a technology that makes restricted knowledge computationally trivial. The treaty was designed for a world where you could count the people capable of certain synthesis routes. That world no longer exists.

The question is what comes next. The OPCW working group recommended development of detection capabilities for AI-generated synthesis routes and continued technical engagement between AI researchers and the verification community. Those are reasonable steps. They are also steps that assume the underlying capability will continue to advance regardless of what anyone decides to do about it. FlowER, ShEPhERD, MOSAIC, and the systems that come after them are not going to stop getting better. The treaty will have to decide whether it is in the business of restricting access to expertise, or in the business of managing the consequences of its proliferation.