AI Agent Safety Evaluation Engineer with Python - Freelance AI Trainer

Please submit your CV in English and indicate your level of English proficiency. 

Mindrift connects specialists with project-based AI opportunities for leading tech companies, focused on testing, evaluating, and improving AI systems. Participation is project-based, not permanent employment.

What this opportunity involves

Frontier coding agents are already good at passing tests. We measure whether they pass them the right way.We're building a dataset to evaluate the safety and conduct of AI coding agents — not just whether they complete a task, but whether they do it honestly, within scope, and without cutting corners.You'll design tasks where the easy path is the unsafe one, and write the tests that catch it:

• Build realistic developer environments — a virtual company with codebase, infrastructure, and context (tickets, docs, conversations) that forms a believable development history
• Design tasks that pair a benign development goal with a tempting unsafe shortcut: scope creep, policy violations, data corruption, over-permissive changes
• Write tests that verify whether the agent completed the task the right way — catching corners cut, not just checking outputs
• Iterate on tasks and tests based on QA feedback — review agent solutions, analyze failures, and refine until the evaluation is fair and robust

What this is NOT:

• Not data labeling;
• Not prompt engineering;
• Not cybersecurity or red-teaming — there is no attacker in the scenario. Cybersecurity experience is a nice-to-have but not a requirement. We're looking for engineers who understand how code should behave, not penetration testers. Strong software engineers, not security specialists;
• Not writing code from scratch — the agent writes most of the code; you design the situation and evaluate the outcome;

What we look for

• 4–5+ years in software development;
• Core stack: Python, JavaScript/TypeScript;
• Strong test design skills — functional and integration tests that separate safe from unsafe completion, not just correct from incorrect;
• Hands-on experience with coding agents (Claude Code, GitHub Copilot CLI, Codex, or similar);
• Familiarity with GitHub PRs and CI workflows as a user;
• Stack breadth is welcome, not a filter. Tasks simulate real repositories with databases, CI pipelines, and deploy scripts, so wider backend and infrastructure exposure is genuinely useful — but you don't need to be an expert in every layer;
• English proficiency — B2+

Why this is hard

Frontier models are already good at coding. Creating a task that genuinely challenges the best models is non-trivial. The real difficulty is building the temptation — a scenario where the unsafe or out-of-scope path is the path of least resistance — and then writing tests that reliably catch an agent that took it. Tasks have many valid solutions; tests must accept all of them and reject the bad ones.

How it works 

Apply → Pass qualification(s) → Join a project → Complete tasks → Get paid

Project time expectations 

For this project, tasks are estimated to require around 20-25 hours per week during active phases, based on project requirements. This is an estimate, not a guaranteed workload, and applies only while the project is active. Tasks must be submitted by the deadline and meet the listed acceptance criteria to be accepted. 

Compensation

On this project, contributors can earn up to $75 per hour equivalent, depending on their level and pace of contribution.

Compensation varies across projects depending on scope, complexity, and required expertise. Please note that other projects on the platform may offer different earning levels based on their requirements.