Nuclear Security Engineer for AI Model Red-Teaming

SaidGig; New York (NY); Remote; Full time

Job summary

SaidGig is seeking a remote, per-task analyst to help evaluate how advanced AI systems handle sensitive nuclear materials and safeguards questions. You will classify prompts as benign, dual-use, or adversarial and craft reference answers that explain the reasoning behind the proper response.

Work is remote, per-task, with briefings in advance and the option to pause or step away as needed. Compensation is per task and reflects the complexity of each assignment.

Qualifications

• Experience in nuclear material control and accountability programme design, operations, or assessment.
• Experience in physical protection system engineering and performance testing.
• Experience with vulnerability assessments or adversary path analysis for nuclear facilities.
• Experience in insider-threat or human-reliability programmes.
• Experience in nuclear security policy, training, or international assistance.
• Red-teaming experience is preferred.
• Strong technical writing ability with clear rationales for non-specialist readers.
• Published research or expert-witness experience valued; include a writing sample or link.

Responsibilities

• Create challenging single-turn nuclear security prompts classified as benign, dual-use, or adversarial.
• Evaluate model responses against a policy standard and determine correctness of handling.
• Develop reference answers with appropriate justification and technical reasoning.
• Apply professional judgment to scenarios where routine material-accounting or protection work could reveal misuse information.

Skills

Nuclear material control

Physical protection engineering

Vulnerability assessment

Insider-threat programs

Nuclear security policy

Red-teaming

Technical writing

Writing samples

Job description

Role Overview

Help evaluate how advanced AI systems handle sensitive nuclear materials and safeguards questions. This remote, per-task role focuses on distinguishing legitimate professional requests from dual-use or adversarial requests, so models can provide useful technical support while appropriately declining unsafe assistance.

Key Responsibilities

• Create challenging single-turn nuclear security prompts classified as benign, dual-use, or adversarial.
• Evaluate model responses against a defined policy standard and determine whether each response was handled correctly.
• Develop reference answers that show the appropriate response and explain the technical reasoning behind it.
• Apply professional judgment to scenarios where routine material-accounting, physical-protection, or vulnerability-assessment work could also reveal misuse-relevant information.

Qualifications

• Experience in nuclear material control and accountability programme design, operations, or assessment at a national laboratory or licensee.
• Experience in physical protection system engineering, including sensors, barriers, delay, response forces, or performance testing.
• Experience with vulnerability assessments or adversary path analysis for nuclear facilities.
• Experience in insider-threat or human-reliability programmes.
• Experience in nuclear security policy, training, or international assistance programmes.
• Red-teaming experience is preferred.
• Strong technical writing ability, with the ability to provide clear written rationales for non-specialist readers.
• Published research, prior technical writing, or expert-witness experience is valued; include a writing sample or link with your application.

Work Terms

Remote, per-task engagement.

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

Compensation

65 to 75 per task.

Eligibility

Work must not involve or rely on classified information, export-controlled information, material covered by an NDA, or material subject to prepublication review obligations. Candidates with these obligations may still be considered. Disclose them in your application so the work can be scoped appropriately.