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 material-balance calculation that closes an inspector's account can be used to hide what is missing from it; the same fuel-cycle step that supplies a power reactor is the one a proliferation pathway runs through. A model that refuses a routine accountancy 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 verified nuclear material rather than read about it — inspection regimes, accountancy, and the judgment calls that come with an unresolved discrepancy. What matters most is being able to tell a routine professional question from one that's fishing for something else: the same material-balance figure that closes an account can be used to conceal a diversion. 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):