Radiologicals Safety for Redteaming

AI safety evaluation

A specialist panel evaluates AI responses involving radiological safety, source security and response. The objective is appropriate handling of technical requests: complete answers for legitimate uses and refusal where genuine danger is present.

Candidates should have been accountable for radioactive materials through custody, licensing, security or emergency response. Working near sources alone is not the intended background. Professional reviewers must distinguish ordinary dose, shielding and isotope questions from attempts to obtain harmful information while allowing legitimate medical and industrial uses to be addressed fully.

Deliverables

• Develop challenging, single turn prompts in your specialty and label each benign, dual use or adversarial.
• Assess the model's replies against a specified policy and determine whether its handling was correct.
• Produce the reference response, with the technical reasoning that supports the expected answer.

Relevant professional backgrounds

Prior adversarial AI testing experience is particularly useful. Applicable experience includes:

• Broad scope Radiation Safety Officer responsibility as the licence holder for acquisition, inventory, transfer and disposal.
• Health physics at a DOE site or national laboratory, covering surveys, assay, contamination control and dose planning in high hazard facilities.
• Assessment of security vulnerabilities affecting Category 1 and 2 sources, including theft and unauthorised movement risks.
• Radiological emergency response through FRMAC or CBRN work, including dispersal and contamination modelling.
• Nuclear medicine physics or cyclotron radiation safety with responsibility for producing and handling medical isotopes.

Writing and working boundaries

Every assessment needs a written explanation understandable to a reader without specialist training. Technical writing, research publications or expert witness work can demonstrate this ability; include a writing sample or a link to one.

The assignment involves sustained reading and writing about misuse scenarios. Experts receive an advance briefing and may pause or leave the work without penalty.

Do not draw on classified material, export controlled information, NDA protected content or material subject to prepublication review. Existing obligations do not necessarily prevent participation: disclose them in your application so the scope can be adjusted.