# Research Scientist Intern - PAR Safety Evals and Red Teaming at Meta

- Company: [Meta](https://meta.com)
- Location: Menlo Park, CA
- Type: Internship · Internship
- Posted: 2026-11-03
- Apply: https://www.metacareers.com/jobs/1903232531024141
- Page: https://newgrad.ai/jobs/research-scientist-intern-par-safety-evals-and-red-teaming-94f5d650

## Summary by newgrad.ai

What you'll do:

- Assess LLM-based judges for safety policy violations and consistency.
- Investigate prompt-injection risks within agent-based tool environments.
- Develop automated red teaming methods to identify safety failures.
- Utilize synthetic data to enhance benchmark design and coverage.

What they're looking for:

- Interest in agent safety, synthetic data, and automated red teaming.
- Experience with LLM-as-a-Judge, benchmark design, and eval harnesses.
- Availability for a 16-week internship duration.

## Posting

Meta was built to help people connect and share, and over the last decade our tools have played a critical part in changing how people around the world communicate with one another. In the pursuit of personal superintelligence, Meta's AI systems are becoming more capable, and evaluating their safety is critical.

Meta is seeking Research Interns to join our Safety Evals and Red Teaming team within the Products and Applied Research org. We are seeking individuals passionate in areas such as Agent Safety, Benchmarks and Evaluations, Synthetic Data Generation, Automated Red teaming, LLM-as-a-Judge, benchmark design, eval harnesses, prompt injection, agent environments.

Our internships are 16 weeks long and we have various start dates throughout the year.

### What You Could Work On:

Depending on your interests and the team’s priorities, a project could address questions such as:

- How reliably do LLM-based judges identify safety-policy violations across different tasks, prompts, and adversarial conditions? How should we measure their errors and consistency?
- How should we evaluate prompt-injection risks when an agent interacts with tools and untrusted information? Which environments and attack scenarios expose failures that simpler tests miss?
- Which automated red teaming methods uncover meaningful safety failures, and how well do those failures generalize beyond the prompts or environments used to find them?
- How can we use synthetic data and benchmark design to improve coverage while checking that scores reflect the behavior we intend to measure?

From Meta's public posting, laid out for reading by newgrad.ai with the wording unchanged. Check the original before you apply.

---

More new-grad and intern roles: https://newgrad.ai/llms.txt