About the work
Mercor is assembling a panel of chemistry and chemical safety experts to red-team frontier AI models. The goal is to test whether a model can correctly judge the misuse potential of a technical request — answering legitimate questions fully while refusing genuinely dangerous ones.
You will
- Write challenging single-turn prompts in your domain, labelled across three levels: benign, dual-use, and adversarial.
- Evaluate the responses that come back against a defined policy standard, and judge whether each was handled correctly.
- Write the reference answer — what a correct response looks like, and the technical reasoning for why.
This domain lives on the dual-use line more than most. The same route design that scales a medicine underpins a controlled compound; the same impurity profile that proves a clean process identifies how something was made. A model that refuses a routine scale-up or kinetics question is failing as badly as one that answers a question it shouldn't. Drawing that line reliably is why we need practitioners rather than generalists.
Background we're looking for
We're looking for chemists who have actually run routes and scaled them — bench to pilot to plant — and who know which requests are ordinary process chemistry and which are asking for something else. What matters most is being able to tell a routine professional question from one that's fishing for something else: the same route knowledge that shortens a synthesis serves whoever is shortening it. You will be writing prompts that sit exactly on that line, then judging whether the model held it.
Examples of relevant backgrounds (ideally with red-teaming experience):
- Synthetic route design and development in pharmaceutical, agrochemical or fine chemicals