About the work
Mercor is assembling a panel of radiological 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 isotope underpins cancer therapy and industrial radiography; the same shielding calculation protects a worker and conceals a source. A model that refuses a routine dose 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 how Category 1 and 2 sources actually get stolen, shielded or moved — the security and vulnerability side of a licensed programme, not just its paperwork. What matters most is being able to tell a routine professional question from one that's fishing for something else: the same shielding calculation protects a worker and conceals a source. 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):
- Security or vulnerability assessment of licensed facilities holding Category 1 and 2 sources