LLM Red Teaming and Safety Prompts Intern

Every assistant we run can be talked into something it should not do, and the only real question is whether we find the route first or a user does.

Eight weeks of breaking things

This short internship puts you on the adversarial side. You will systematically attack our candidate assistant and our recruiter agent: prompt injection hidden in uploaded resumes, instructions buried in job descriptions, roleplay framings that coax out the system prompt, and requests that nudge the model toward discriminatory screening advice. Finding a break is the easy half. You then write the guardrail prompt or input filter that closes it, add the case to a permanent regression suite, and confirm the fix has not made the assistant uselessly cautious. That tradeoff between safety and usefulness is the actual work.

Terms, stated plainly

The position is unpaid and runs for two months, with two days a week in the Bengaluru office. You will work directly with the engineer responsible for model safety, your test cases become part of what ships, and you will finish with a body of documented findings that demonstrates a real and rare skill.