Red Team Researcher, Cyber Capabilities and Safeguards
AI research and deployment company; investigations team focused on detecting and mitigating abuse of its products
| Employer | AI research and deployment company; investigations team focused on detecting and mitigating abuse of its products (not named) |
|---|---|
| Location | San Francisco, CA (Bay Area) · Hybrid |
| Type & level | Full-time · Mid-level |
| Pay | USD 198k–320k + equity Source: pay range stated in the source posting |
| Visa path | O-1A fit |
| Role family | AI Red Team / ML Security Researcher |
About the role
This team tests how much real-world help AI models could give to cyber attackers, and whether safeguards still hold against skilled adversaries. The role pairs large-scale evaluation with expert hands-on testing. Part of the time goes to abuse risks in agentic systems and to safety bug bounty reports that need cyber expertise.
What you'll do
- Design and run evaluations of model cyber capabilities and safeguards, covering policy adherence, correct refusal, over-refusal and jailbreak resilience.
- Test by hand what models enable for experienced security practitioners, using task-specific harnesses, scaffolding and multi-step workflows.
- Separate benchmark or policy failures from behavior that creates meaningful real-world risk, weighing feasibility, attacker uplift and reliability.
- Build automated testing infrastructure for repeatable, statistically grounded measurement across models and product surfaces.
- Probe agentic systems for indirect prompt injection, agent hijacking and other ways adversaries can manipulate tool-using systems.
- Turn findings into clear risk assessments and recommendations for security, research, product, policy and engineering partners.
Requirements
- Substantial depth in either cybersecurity (application security, penetration testing, vulnerability research, adversary simulation or red-team operations) or AI model evaluation (evals, agentic harnesses, adversarial test automation).
- Working literacy in the other area, with interest in building further depth.
- Ability to write code and build practical tools for automating experiments, orchestrating models and analyzing results.
- An attacker mindset and interest in failure modes that standard evaluations miss.
- Clear written and verbal communication of technical findings, limitations and risk to varied audiences.
- Experience working across technical and non-technical teams to move from a finding to a decision or mitigation.
Nice to have
Pay & benefits
USD 198k–320k + equity. The posting lists equity alongside base pay and offers relocation assistance to new employees. It does not describe other benefits.
Pay source: pay range stated in the source posting. Market data, not an offer from O1DMatch.
Visa path
People in AI red teaming and security research typically support an O-1A case with evidence such as published vulnerability research or CVEs, conference talks, bug bounty recognition, expert judging or peer review, and high pay for the field. The posting states no citizenship restriction. An employer like this, or a U.S. agent acting for the candidate, can file the petition, and approval depends on the individual's evidence.
How to get matched
- Take the free eligibility checkAn AI review of your background against the eight O-1 criteria, at check.o1dmatch.com.
- Create your profile with code WORKWITHUSIt unlocks the Starter plan at no cost. Tell us your field and the work you do today.
- Upload your evidenceEach document is sorted under the criterion it supports, and your criteria map shows what is still missing.
- Get matchedWe match strong profiles with U.S. employers hiring for roles like this one. Gigs you confirm with verified companies count toward your itinerary, and at a score of 75 we connect you with an immigration attorney.
This listing describes a real U.S. posting so you can see what employers ask for. O1DMatch does not represent that employer and does not hold this opening; applying matches you with employers hiring for this kind of role. No visa outcome is promised.