Nuclear Engineering & Safeguards Experts for Red Team

Nuclear domain AI evaluation

A panel of nuclear specialists tests whether AI models distinguish legitimate technical requests from genuinely dangerous ones. The aim is to preserve complete, useful answers for appropriate questions without allowing harmful requests through.

Evaluation responsibilities

• Write challenging single turn prompts and classify them as benign, dual use or adversarial.
• Assess returned model answers against an established policy standard.
• Prepare reference answers explaining the correct response and its supporting technical reasoning.

The work involves recognising when apparently academic or research oriented language masks a different intent. Reviewers must explain why a response respects or crosses the appropriate boundary.

Relevant professional experience

Fuel cycle depth is expected, together with exposure to material control, accounting or diversion risks. Safeguards and nonproliferation backgrounds are especially valuable, and previous adversarial AI testing experience is useful.

• Safeguards analysis or IAEA/Euratom inspection involving material accountancy, inspection regimes and the Additional Protocol.
• Nuclear fuel cycle engineering covering enrichment, reprocessing and criticality across the cycle.
• National laboratory nuclear security engineering, including MPC&A, insider threats and physical protection design.
• Nuclear forensics involving radioxenon monitoring, attribution, detection and characterisation.
• Nonproliferation analysis involving export controls, NSG trigger lists and acquisition or smuggling pathways.

Writing and information boundaries

Every judgment needs a written rationale understandable to nonspecialists. Include evidence of technical writing, published research or expert witness work through a sample or link.

The work entails sustained exposure to misuse scenarios in the field. Experts are briefed beforehand and may pause or step away without penalty.

Classified, export controlled, NDA protected or prepublication review material must not inform the work. Disclose any such obligations when applying so the assignment can be scoped appropriately; those obligations do not automatically rule out participation.