Nuclear Engineering and Safeguards Expert AI Red Teaming

Help test frontier AI models’ ability to distinguish legitimate nuclear technical questions from requests with meaningful misuse potential. This writing-focused role uses your nuclear, safeguards, or nonproliferation expertise to create realistic boundary-testing scenarios and explain what a safe, technically sound response should do.

• Develop challenging single-turn prompts in your specialty and label them as benign, dual-use, or adversarial.
• Assess model responses against an established policy standard and determine whether each response was handled appropriately.
• Prepare reference answers that describe the correct response and the technical reasoning behind it.
• Write clear rationales that non-specialists can understand, including why a response is acceptable or unacceptable.

• Deep fuel-cycle knowledge and familiarity with how nuclear material is controlled, accounted for, or diverted.
• Safeguards or nonproliferation experience is especially relevant.
• Relevant backgrounds may include safeguards analysis or IAEA/Euratom inspection work involving material accountancy, inspection regimes, or the Additional Protocol; nuclear fuel-cycle engineering involving enrichment, reprocessing, or criticality; nuclear security engineering involving material protection, control and accounting, insider threats, or physical-protection design; nuclear forensics involving radioxenon monitoring, attribution, detection, or characterization; or nonproliferation analysis involving export controls, NSG trigger lists, acquisition, or smuggling pathways.
• Red-teaming experience is preferred.
• Strong technical writing experience, published research, or expert-witness experience is valued.
• Include a writing sample or link with your application.

Remote, per-task independent contractor engagement.

The work involves sustained reading and writing about nuclear misuse scenarios. You will receive advance briefing and may pause or step away at any time without penalty.

Pay is 65 to 75 per task.

Do not use or draw on classified, export-controlled, NDA-protected, or prepublication-review information. If you have such obligations, disclose them in your application so the work can be scoped appropriately.