Cybersecurity Specialist (AI Safety)

You will build complete, sandboxed evaluation tasks for AI agents. Each task includes an instruction for the agent, a containerized target system, a hidden deterministic verifier that checks the result, and a reference solution.

The work focuses on AI red teaming rather than traditional vulnerability assessment and penetration testing. You will test how agents respond to prompt injection, jailbreaks, tool-call abuse, unsafe autonomous actions, scope escape, and reward hacking.

• Author security scenarios and agent-facing instructions from scratch.
• Build Dockerized target environments and write oracle solutions.
• Create deterministic verifiers for checking agent behavior.
• Judge multi-step agent trajectories and identify unsafe or out-of-scope actions.
• Test prompt injection, jailbreaks, tool misuse, and reward-hacking risks.

What it pays and takes

This is a remote freelance contractor engagement for part-time work. The role is open worldwide and is intended for contributors with intermediate experience in AI safety evaluation or related red-team work.

• Pay: $300 per accepted task, paid after quality control.
• Time: Less than 20 hours per week.
• Location: Remote and open worldwide.
• Experience: Hands-on LLM or AI-agent red teaming, jailbreak or prompt-injection research, or AI safety evaluation.
• Technical skills: Docker, Git, and Python.
• Required ability: Write deterministic verifiers and evaluate multi-step agent trajectories.
• Useful background: Published red-team reports, adversarial machine-learning research, or reinforcement-learning environment experience.

How it works

Apply on OpenTrain. The employer reviews applications there.

• Selection includes a short take-home assignment and one feedback round.

About AI training work

AI training is the human work used to build and evaluate artificial intelligence, including testing model responses and reviewing how agents behave. People with security and technical expertise are needed to find failure modes and help make AI systems safer.