Adversarial Prompt Expert

As an Adversarial Prompt Expert, you will test large language models to find unsafe, unexpected, or otherwise problematic behavior. The work is project-based, remote, largely asynchronous, and includes occasional recommended synchronous meetings.

You will document your attempts and findings clearly so they can support stronger AI system defenses, safety work, and engineering decisions.

• Probe multiple open-source and closed-source language models.
• Create adversarial prompts, jailbreak attempts, and edge-case test scenarios across different topics.
• Review model responses for failures, harmful outputs, and weaknesses in safeguards.
• Try different approaches persistently and record attempts and outcomes systematically.
• Research relevant topics with AI assistance and contribute documented insights.

What it pays and takes

Project placement depends on available work.

• Pay: Up to $45 to $65 per hour, depending on prior adversarial testing experience.
• Schedule: 20 or more hours per week, with flexible scheduling.
• Location: Open to candidates based in the United States.
• Hands-on experience using multiple open-source and closed-source language models.
• Strong prompt engineering skills, including evasion techniques and jailbreaking.
• An adversarial or security mindset, with the ability to think like an attacker and test edge cases.
• Persistence, creativity, careful documentation, and clear written communication.
• Ethical judgment and the ability to handle sensitive content responsibly.
• Helpful background includes red teaming, offensive security, model evaluation, or comparable experience probing technical systems.

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 writing prompts, rating model responses, and testing how models behave. People with experience in areas such as security, model evaluation, or specialized testing can help identify difficult failures and improve how AI systems respond.