About the work
Mercor is assembling a panel of energetic materials and propulsion 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 assessment that tells a technician how to approach a device tells someone else how it was meant to work; the same compatibility data that keeps a magazine safe describes what stays stable elsewhere. A model that refuses a routine storage or render-safe 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 public safety professionals who have assessed real devices — the render-safe and post-incident side, held under certification and command accountability. What matters most is being able to tell a routine professional question from one that's fishing for something else: the same assessment that makes a device safe explains how it was built. 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 bomb technician or hazardous device operator — device assessment and render-safe procedures