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 dispersion model that sites a plant and sets an evacuation radius also predicts an exposure; the same hazard analysis that finds a failure path is a list of failure paths. A model that refuses a routine exposure-limit 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 people who have assessed real exposures and real process hazards, and signed off on the controls. What matters most is being able to tell a routine professional question from one that's fishing for something else: the same release model that protects a community estimates a consequence. 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):
- Certified industrial hygienist — exposure assessment, sampling strategy, control banding
- Process hazard analysis — HAZOP, LOPA, what-if studies, management of change