Nuclear Security Engineer

AI safety evaluation

Contribute practical nuclear security expertise to an AI safety testing panel. The aim is to assess whether models answer legitimate technical questions accurately while recognizing and declining requests that could enable harm.

The work requires professional judgment about context and intent. Both unjustified refusals of routine professional questions and responses that provide inappropriate assistance count as failures.

Responsibilities

• Write challenging single-turn prompts labeled as benign, dual-use or adversarial.

• Evaluate model responses against the defined policy standard and record whether each was handled correctly.

• Write a reference answer showing the appropriate response and the technical rationale supporting it.

Relevant professional background

The role calls for people who have been responsible for keeping nuclear material where it belongs — material control and accountability, physical protection, and the assessments that test both. Professional judgment must distinguish legitimate requests from potential misuse: the same vulnerability assessment that closes a gap also describes it.

Create prompts that test that distinction and assess whether the response respects it. Relevant backgrounds include the following; prior AI safety testing is helpful:

• Material control and accountability (MC&A) programme design, operation or assessment at a national laboratory or licensee

• Physical protection system engineering — sensors, barriers, delay, response force, performance testing

• Vulnerability assessment and adversary path analysis for nuclear facilities

• Insider threat and human reliability programme work

• Nuclear security policy, training or international assistance programmes

Writing and working conditions

Every evaluation needs a clear written rationale that a non-specialist can understand. Technical writing, published research or expert witness experience is valuable; a writing sample can demonstrate that background.

Assignments involve sustained reading and writing about misuse scenarios. Experts receive a briefing in advance and may pause or step away without penalty.

Work must exclude classified or export-controlled information and material restricted by an NDA or prepublication review obligation. Existing obligations should be disclosed so assignments can be scoped appropriately.