RL Environment Engineer for AI Safety Coding Agents

You will create complete reinforcement learning environments for realistic coding tasks. These environments help evaluate how AI coding agents solve problems, use tools, handle security issues, and respond to failures.

You will also review agent reasoning traces, find failure causes, and improve scenario packages using structured quality feedback. Work includes secure coding, prompt injection testing, AI red teaming, and tool misuse testing.

• Design coding scenarios, fixtures, and repositories.
• Write production-quality Python code, reference solutions, pytest harnesses, and deterministic verifiers.
• Build containerized, offline environments with Docker.
• Review agent trajectories and diagnose why tasks failed.
• Test vulnerabilities and apply appropriate remediation.
• Revise benchmark packages after quality review.

What it pays and takes

This is a remote, part-time contractor role for experienced software, security, or machine learning engineers. Tasks are paid after they pass quality review and can be completed within flexible deadlines.

• Pay: $200 fixed price per task.
• Time: Less than 20 hours per week.
• Location: Open worldwide and fully remote.
• Experience: Intermediate level; experience with AI evaluation, benchmark authoring, red teaming, application security, penetration testing, CTFs, or coding agents is highly valued.
• Required skills: Strong Python, Docker fluency, production-quality coding, pytest, Linux and shell, and the ability to build offline environments.
• You should understand software vulnerabilities and remediation, review agent reasoning traces, diagnose failures, and use an adversarial but fair approach.

How it works

Apply on OpenTrain. The employer reviews applications there.

About AI training work

AI training is the human work behind modern artificial intelligence, including writing code, creating evaluations, testing model behavior, and reviewing AI outputs. Experienced contributors are paid for applying technical judgment to help make AI systems more reliable and safer.