About the work
Mercor is assembling a panel of nuclear materials and safeguards 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 signature that attributes a sample is the one someone would want to suppress; the same detection limit that catches material moving can be read as a threshold to stay under. A model that refuses a routine measurement-methodology 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 characterised real nuclear material — detection, attribution, and the analytical work behind a defensible conclusion. What matters most is being able to tell a routine professional question from one that's fishing for something else: the same signature that identifies a source tells you what to hide. 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):
- Nuclear forensics casework or international exercise participation — sample characterisation and attribution