Research Scientist, Open Benchmarks Red Team

Open Benchmarks Grants is Snorkel’s $30M commitment to fund and support researchers and independent builders developing new AI evaluations. We support the teams behind Terminal-Bench, OSWorld 2.0, and Senior SWE-bench, and run a Research Fellows program for PhD students and postdocs. The program funds work in areas including physical AI, scientific discovery, cybersecurity, safety, and alignment, and helps teams test and strengthen their benchmarks before and after public release.

Snorkel AI is hiring researchers to support the Open Benchmarks Red Team program.

Frontier AI benchmarks can contain grading failures, exploitable shortcuts, and task errors that make results unreliable. We’re looking for a Research Scientist / Research Engineer to identify benchmark weaknesses, help develop and validate fixes, and create reusable methods to improve benchmark quality across tasks and datasets.

This is a hands-on research role combining AI evaluation, adversarial testing, practical benchmark repairs, and close collaboration with benchmark authors. You'll share findings with benchmark authors, work together on fixes, and help get improved benchmark versions adopted.

• Red-team frontier benchmarks: Identify and reproduce issues such as reward hacking, grader failures, task ambiguity, shortcuts, and environment vulnerabilities.
• Develop benchmark repairs: Propose next steps for validated issues, implement selected fixes with benchmark authors, and verify fixes without breaking valid solutions.
• Develop reusable methods: Reuse existing tools where possible and develop new checks to detect recurring benchmark weaknesses.
• Collaborate with benchmark authors: Review findings with benchmark authors before publication, collaborate on repairs, and support official improved releases.
• Drive research impact: Contribute to technical reports, open-source tools, publications, and new approaches to benchmark evaluation and improvement.

• Strong research or engineering background in machine learning, LLMs, AI agents, and evaluation.
• Strong Python programming skills and experience building, running, and debugging AI agents or evaluation systems.
• Ability to identify weaknesses in existing evaluation methods, develop new approaches, and rigorously test whether they work.
• Experience designing rigorous experiments, analyzing model behavior, and investigating unexpected failures.
• Familiarity with LLM/agent benchmarks, evaluation frameworks, or automated grading systems.
• Experience independently driving research and engineering projects from problem identification through implementation and validation.
• Strong communication and collaboration skills.

• Experience with adversarial testing, red teaming, reward hacking, or specification gaming.
• Experience developing or evaluating benchmarks for coding agents, computer use, professional workflows, etc.
• Familiarity with RL environments, agent trajectories, verifiers, and evaluation infrastructure.
• Open-source contributions or publications in AI evaluation, robustness, or agentic systems.
• PhD in Computer Science, Machine Learning, or a related field, or equivalent research/industry experience.