AI Red Teamer, CBRNE

Handshake•Seattle, WA

2d•$65 - $158

About The Position

As a CBRNE Red Teamer, you will evaluate whether AI models appropriately handle queries related to chemical, biological, radiological, nuclear, and explosive threats. Your job is to probe models for dangerous knowledge gaps in their safety guardrails, testing whether they can be manipulated into providing meaningful uplift toward the creation, acquisition, or deployment of weapons or hazardous materials. This work sits at the intersection of AI safety and national security. You will use your domain expertise to craft realistic, technically grounded adversarial scenarios that stress-test model defenses far beyond what a non-expert could attempt. The goal is not to generate harmful content, but to find and document the places where models fail to refuse, hedge, or redirect appropriately so that labs can fix them before those failures reach the real world. This role requires deep subject matter expertise in at least one CBRNE domain, strong ethical judgment, and the ability to think like a sophisticated threat actor while operating within a structured evaluation framework.

Requirements

Graduate-level education or equivalent professional experience in a relevant CBRNE field (chemistry, biochemistry, microbiology, virology, nuclear physics, radiochemistry, materials science, munitions/ordnance, chemical engineering, or closely related disciplines)
Ability to evaluate the technical accuracy and real-world consequence of model outputs in your domain
Understanding of dual-use research concerns and the distinction between open-source knowledge and operationally significant uplift
Strong hands-on experience using multiple LLMs (ChatGPT, Claude, Gemini, open-source models, etc.)
Creative, adversarial problem-solving skills
Clear and precise written communication, including the ability to explain technical risk to non-specialist audiences
Strong ethical judgment and the ability to separate adversarial thinking from personal values
Self-directed, collaborative, and comfortable in feedback-heavy environments

Nice To Haves

Active or prior security clearance (Secret, Top Secret, or SCI)
Experience in threat assessment, WMD analysis, intelligence analysis, or arms control verification
Background in biosafety/biosecurity, chemical safety, nuclear nonproliferation, or explosive ordnance disposal
Familiarity with relevant regulatory frameworks (CWC, BWC, IAEA safeguards, ATF regulations, Export Administration Regulations)
Experience in red teaming, penetration testing, or structured adversarial evaluation in any context
Familiarity with Python or scripting languages, LLM APIs, or evaluation tooling
Published research or professional presentations in a relevant CBRNE domain
Prior work in trust and safety, content moderation, or AI evaluation

Responsibilities

Design technically grounded adversarial prompts that test whether models provide meaningful uplift toward CBRNE threats
Evaluate model outputs for technical accuracy, assessing whether responses contain genuinely dangerous information versus superficial or publicly available knowledge
Probe dual-use knowledge boundaries, testing how models handle queries that blend legitimate scientific, medical, or industrial use cases with potential weapons applications
Test multi-step and multi-turn attack chains that simulate how a motivated actor might extract dangerous information incrementally
Score model responses against structured harm taxonomies and severity rubrics calibrated to real-world risk
Document findings with clear technical reasoning, including what a response gets right, what it gets wrong, and why the failure matters
Identify and articulate the difference between information that is freely available in open literature and information that constitutes genuine uplift beyond baseline
Contribute to the development and refinement of CBRNE-specific evaluation frameworks and threat models
Collaborate with other red teamers, AI researchers, and policy teams to translate findings into actionable model improvements
Stay current on evolving model capabilities, jailbreak techniques, and relevant developments in your domain