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 while 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 held responsibility for radioactive material, not just worked near it; source custody, licensing, or the security and response side. 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, They will be writing prompts that sit exactly on that line, then judging whether the model held it. Examples of relevant backgrounds (Ideally w/Red-teaming experience):
Radiation Safety Officer (broad-scope licensee) — holds the licence of record; acquisition, inventory, transfer and disposal.
Health Physicist (DOE site or national lab) — survey, assay, contamination control and dose planning in high-hazard facilities.