The work
You will create training and evaluation tasks for AI coding models. The work turns real software vulnerabilities into audit and patch exercises that test whether a model can identify, understand, reproduce, and securely fix a security issue.
You will work with vulnerabilities from CVEs, security advisories, bug reports, and open-source repositories. You will also review AI-generated patches and reasoning against technical rubrics.
- Classify vulnerabilities using the most specific applicable CWE and relevant OWASP Top 10 categories.
- Reproduce vulnerabilities and write reliable proof-of-vulnerability tests.
- Build Harbor-format audit and patch tasks in Python, JavaScript or TypeScript, Java, C or C++, C#, Go, Rust, Ruby, PHP, Kotlin, and Swift.
- Write realistic task instructions and create reproducible Docker environments with pinned dependencies.
- Harden verifiers against gaming and test manipulation.
- Review AI-generated code patches and explanations using technical evaluation rubrics.
What it takes
This is a remote contractor role for at least 20 hours per week. The work requires professional software engineering experience, strong software security knowledge, and the ability to work independently in a research-focused setting.
- At least five years of professional software engineering experience.
- Expert proficiency in at least one supported programming language.
- Professional fluency in English.
- Strong knowledge of CWE, CVE, and the OWASP Top 10.
- Ability to reproduce vulnerabilities and write proof-of-vulnerability tests.
- Experience with sanitizers, fuzzers, and static analysis tools such as CodeQL or Semgrep.
- Strong Git, Linux, Bash, Docker, technical writing, and documentation skills.
- Open to candidates worldwide.
- Experience building security benchmarks, analyzing open-source vulnerabilities, creating automated tests, or evaluating code patches is helpful.
How it works
Apply on OpenTrain with your resume, then complete the application on the hiring site.
About AI training work
AI training is the human work behind systems that learn from examples, including code tasks, model responses, and evaluations. Experienced contributors are needed to create accurate examples and reviews that help AI systems produce safer, more useful results.