AI Red Teamer, CBRNE (Remote)

Written by the T&S Network from Handshake's posting.

AI red teaming for chemical, biological, radiological, nuclear, and explosive threats. The work probes whether AI models refuse, hedge, or redirect on CBRNE queries, so labs can fix failures before they reach the real world. It sits at the intersection of AI safety and national security, on frontier AI lab model evaluations.

What you'd do

• Design technically grounded adversarial prompts testing whether models give meaningful uplift toward CBRNE threats.
• Judge model outputs for technical accuracy and whether they contain dangerous information or only public knowledge.
• Probe dual-use queries that mix scientific, medical, or industrial use with possible weapons applications.
• Run multi-step and multi-turn attack chains that mimic incremental extraction by a motivated actor.
• Rate each response on a harm taxonomy and severity scale calibrated to actual risk.
• Document what each response gets right and wrong, and why the failure matters.
• Contribute to CBRNE-specific evaluation frameworks and threat models.

What it asks for

• Graduate degree or equivalent work background in a field like chemistry, microbiology, nuclear physics, or ordnance.
• Advanced working knowledge of one CBRNE area.
• Grasp of dual-use risk and how to tell public information from real capability gain.
• Strong hands-on experience using multiple LLMs.
• Clear written communication that explains technical risk to non-specialists.
• Current or past U.S. security clearance at Secret, Top Secret, or SCI level (preferred)

Worth knowing

• Regular engagement with sensitive content involving weapons of mass destruction, biological agents, radiological and nuclear materials, and explosive devices.
• Work is conducted under strict ethical guidelines and operational security protocols.
• Candidates must be able to engage with this material professionally and sustainably.

Terms in this role

Red teaming.

Testing a system, often an AI model, by deliberately trying to make it fail or misbehave, so problems are found and fixed first.

Jailbreak.

A prompt written to get an AI model to ignore its safety rules.

Our best summary of the posting, and it can be wrong. The employer's posting is what counts: read it in full on Handshakenorth_east before you apply. If we got something wrong, report this listing.