Radiation Safety Officer - Remote AI Red Team
Listing checked September 15, 2026 · pay as published by Mercor
Overview
A panel of radiological safety experts will red-team frontier AI models. The core task is to test whether a model can spot misuse potential in technical requests while still answering legitimate questions. You write single-turn prompts across benign, dual-use, and adversarial levels, score model replies against a policy standard, and draft reference answers with technical reasoning. The domain sits on a difficult line: the same isotope can support cancer therapy or industrial radiography, and the same shielding calculation can protect a worker or hide a source. The work needs practitioners who have held the licence, because generalists often miss that boundary.
What You'll Do9
- 1Write single-turn prompts in radiation safety that cover benign, dual-use, and adversarial cases.
- 2Label each prompt by risk level and explain the technical boundary it tests.
- 3Review model responses against a written policy standard and decide if the reply meets that standard.
- 4Draft the reference answer for every prompt, with reasoning a non-specialist can follow.
- 5Separate routine professional questions from requests that seek harmful operational detail.
- 6Use licence-level knowledge of acquisition, source inventory, transfer, and disposal to shape test cases.
- 7Flag over-refusals and under-refusals so the evaluation covers both sides.
- 8Provide written rationale for each judgment, because the work demands sustained writing.
- 9Pause or step away from misuse scenarios when needed, without penalty.
Requirements10
- 1Named Radiation Safety Officer or person of record on a broad-scope licence.
- 2Hands-on work with source inventory, acquisition, transfer, disposal, and regulatory compliance.
- 3Experience at a university, hospital, research institute, industrial site, or irradiator facility.
- 4Familiarity with mixed radionuclide programs and sealed sources, including Category 1 and 2 sources.
- 5Knowledge of NRC, Agreement State, or national regulator submissions, amendments, and inspections.
- 6Background in radiography, gauging, sterilisation, ALARA programs, audits, training, or incident reporting.
- 7Ability to tell a routine professional question from one that probes for misuse.
- 8Strong technical writing, published research, or expert witness experience. Include a sample or link.
- 9Red-teaming experience is a plus, not a hard requirement.
- 10No use of classified or export-controlled information, NDA material, or prepublication review content. Disclose any such obligations in the application.
Who Should Apply
The ideal candidate has held the licence as the person of record, managed a broad-scope source program, and written technical justifications that outsiders can follow. That person knows the dual-use line from practice: the same calculation that protects a worker can conceal a source, and the same isotope can treat cancer or support industrial radiography. Prior red-teaming, expert witness work, or published research helps your application stand out. The role fits less well for generalists who have worked near radiation safety but never owned compliance decisions, or for anyone who cannot write a clear rationale for each judgment. Applications lose points when they skip a writing sample, do not name the licence type, or omit NDA, export-control, or prepublication obligations.
Salary Insight
The posting lists a one-time payment of $75 for this task-based engagement. That figure fits a small contribution rather than a full retainer, so ask how many prompts, evaluations, and reference answers the payment covers before you accept.
Location
Compensation
$75 fixed
Required Skills
Application Tip
Lead with your named RSO history on a broad-scope licence, then attach a short writing sample that explains a dual-use boundary in plain language. Name the sealed source categories, regulators, and inspections you have handled so the reviewer can map your fit at a glance.
See NearSkill jobs more often in your search
Application & verification flow
1Instant rubric match
Your resume is scanned against this role’s requirements to check qualification fit.
2Screened before the employer sees it
Only profiles that clear screening are passed on.
3Outcome by email
We notify you at the address on your resume once the screening is reviewed.
Similar open positions
Explore active roles that match your skills and interests.
Mercor
VerifiedRadiological Safety AI Red Team Expert (Remote)
Neon seeks radiological safety practitioners to red-team frontier AI models on a task-based engagement. You will craft single-turn prompts that sit on the dual-use line, label each one as benign, dual-use, or adversarial, then judge how the model handles the request against a set policy. The work matters because a useful model must answer routine dose, shielding, or isotope questions while refusing requests that seek misuse details. Your domain knowledge decides where that boundary falls, and your writing makes the call clear to people outside radiological safety.
Mercor
VerifiedRadiological Source Security Specialist - Remote
Mercor needs radiological safety practitioners to red-team frontier AI models. You will craft single-turn prompts that test whether a model can separate legitimate technical questions from dangerous misuse, with each prompt tagged benign, dual-use, or adversarial. The work also calls for reviewing model responses against a policy standard and writing reference answers that show the correct call, plus the reasoning behind it. The domain sits on the dual-use line: the same isotope can support cancer therapy or industrial radiography, and the same shielding math can protect a worker or hide a source.
Mercor
VerifiedRadiological Emergency Response Specialist - Remote
This project brings together radiological safety experts to red-team frontier AI models. The panel tests whether a model can judge the misuse potential of a technical request, answering legitimate questions while refusing dangerous ones. Your domain sits on the dual-use line: the same isotope supports cancer therapy and industrial radiography, and the same shielding math protects a worker or hides a source. A model that refuses a routine dose question fails just as much as one that answers a harmful query. Radiological safety practitioners, not generalists, draw that line.
Mercor
VerifiedHealth Physicist - Remote AI Red-Teaming Contract
A project assembles a panel of radiological safety experts to red-team frontier AI models. The core question: can a model judge the misuse potential of a technical request, answer legitimate questions, and refuse requests that pose real harm; a refusal of a routine dose question counts as a failure, just as an answer that should have been refused does. Experts write challenging single-turn prompts in health physics, label each as benign, dual-use, or adversarial, then assess model replies against a defined policy standard. The work also requires reference answers that show the correct response and the technical reasoning behind it. The domain sits on the dual-use line: the same isotope supports cancer therapy and industrial radiography, and the same shielding calculation protects a worker or conceals a source.
Mercor
VerifiedNuclear Medicine Physicist - Cyclotron RSO Remote
Mercor needs radiological safety practitioners to red-team frontier AI models on questions that sit along the dual-use line. You will craft single-turn prompts in your specialty, tag each as benign, dual-use, or adversarial, and score model replies against a fixed policy. The work hinges on distinctions such as I-131 therapy dosing, Lu-177 and Y-90 handling, or PET isotope production in hot cells. You then write the reference answer and the technical reasoning a correct response should contain. A model that blocks a routine dose question fails just as much as one that answers a dangerous one.
Mercor
VerifiedNuclear Security Engineer (Remote, Task-Based)
A panel of nuclear materials and safeguards specialists will red-team frontier AI models for Mercor, testing how well a model judges misuse potential. You write single-turn prompts at three levels: benign, dual-use, and adversarial, then judge model replies against a policy standard and draft the reference answer with technical reasoning. The work sits on the dual-use line where a material-balance calculation can close an inspector's account or hide a gap. Expect sustained reading and writing about misuse scenarios, with briefings and freedom to pause without penalty.