Artificial Intelligence Researcher (Cybersecurity) - PhD

Stealth AI Startup; San Francisco (CA); On-site; USD 120,000 - 180,000; Full time

Job summary

Stealth AI Startup is hiring PhD graduates to join our on-site cybersecurity team in San Francisco or Sunnyvale, CA, launching full-time around December 2026. You will craft end-to-end environments where AI agents can investigate, fix vulnerabilities in real codebases, and be measured for reliability and security.

You will own the project from security scenario design through runnable environment, instructions, and verifier, contributing to benchmarks and research while learning our AI tooling

Qualifications

• PhD in related technical field is required.
• Hands-on depth in cybersecurity such as app security, vulnerability research, or incident response.
• Ability to read unfamiliar code, reproduce issues, and craft fixes.
• Comfort with Linux, Git, containers, debugging, and automated testing.
• Curiosity about applying cybersecurity expertise to AI; prior LLM experience not required.

Responsibilities

• Build cybersecurity environments with real repositories, apps, services, logs and access controls.
• Design scenarios around vulnerability discovery and remediation, API security, auth, incident investigation and hardening.
• Develop reference solutions and verifiers to prove fixes restore functionality and block attacks.
• Test graders by creating incomplete fixes and hidden data to ensure robust evaluation.
• Run and analyze agents, explain failure points, and inform task coverage improvements.
• Collaborate with researchers to turn environments into benchmarks and training data.

Skills

Linux

Git

Containers

Debugging

Automated testing

Education

PhD in related field

Tools

Git

Docker

Kubernetes

CI/CD

Job description

This role is for PhD New Grads who are available to start full time around December 2026, in person/onsite either San Francisco or Sunnyvale CA.

We build environments, tasks, and evaluations that help frontier AI models improve at real work. We are growing our cybersecurity team and looking for someone who can turn practical security problems into environments where AI agents can investigate, act, and learn.

Agents must find and fix vulnerabilities in real codebases, with checks that confirm the vulnerability is resolved and existing functionality still works. You will help build what comes next: richer cybersecurity environments, realistic tasks, and reliable ways to measure whether agents succeed.

You will own work from the initial security scenario through the runnable environment, task instructions, reference solution, and verifier. We are looking for hands-on security knowledge, strong programming skills, and curiosity about how AI agents fail.

Deep cybersecurity expertise is enough to get started—you do not need prior AI or machine-learning experience.

If you know how to investigate vulnerabilities, reason about security failures, and verify that a fix works, we want to hear from you. We will teach you our AI tooling and evaluation workflows.

What you will do

• Build cybersecurity environments. Create isolated, reproducible environments with real repositories, applications, services, logs, and access controls. Make them straightforward to launch, reset, and evaluate at scale.
• Turn security work into tasks. Design scenarios around vulnerability discovery and remediation, application and API security, authentication and authorization, incident investigation, and system hardening. Define what the agent knows, which tools it can use, and what it must accomplish.
• Develop reference solutions and verifiers. Reproduce the underlying issue, implement a valid solution, and write checks that distinguish a real fix from a superficial workaround. Verify that attacks fail after remediation while legitimate behavior continues to work.
• Test the tests. Challenge graders with incomplete fixes, disabled features, hard-coded answers, and other shortcuts. Keep hidden solutions and test data out of the agent's environment, and separate genuine model failures from broken infrastructure.
• Run and analyze agents. Evaluate frontier models, inspect their tool calls and code changes, and explain where their security reasoning or execution breaks down. Use those findings to improve task coverage and difficulty.
• Contribute to benchmarks and research. Work with researchers and engineers to turn strong environments into evaluation suites and training data. Review other contributors' tasks and help document results for future releases.

Must Have

• PhD in related technical field.
• Practical depth in at least one area of cybersecurity, such as application security, vulnerability research, penetration testing, systems security, cloud security, or incident response. Evidence can come from internships, research, open-source work, bug bounties, CTFs, or independent projects.
• The ability to read unfamiliar code, reproduce a security issue, understand its root cause, and implement or assess a fix.
• Comfort with Linux, Git, containers, debugging, and automated testing.
• Clear technical writing and careful judgment about what an evaluation does and does not demonstrate. You can explain why a task is realistic and why its grading is trustworthy.
• Curiosity about applying your cybersecurity expertise to AI, and a willingness to learn how to evaluate tool-using agents. Prior experience with LLMs is not expected.

Nice to have

• Disclosed vulnerabilities, accepted security patches, strong CTF results, or useful security tools and write-ups.
• Experience with fuzzing, static or dynamic analysis, reverse engineering, or building security labs.
• Familiarity with CWE and OWASP classifications and how they relate to concrete software failures.
• Experience building agent evaluations, adversarial tests, prompt-injection defenses, or reinforcement-learning environments.

Examples of what you might build

• A repository audit where an agent must discover and repair an authorization flaw without being told where it is.
• A small service environment where an agent investigates suspicious activity from logs and configuration, then applies and verifies a remediation.
• An LLM application where an agent must repair a prompt-injection or tool-permission weakness while preserving legitimate functionality.