PeopleofSuperintelligence

Lab leader

Ariel Herbert-Voss

Also known as Ari Herbert-Voss

Ariel Herbert-Voss is an AI security researcher and, as of October 2026, the CEO and cofounder of RunSybil. They studied computer science and applied mathematics at Utah before doctoral work at Harvard, where they focused on adversarial machine learning and co-founded DEF CON's AI Village. At OpenAI, they coauthored the GPT-3 few-shot learning paper, which received a 2020 NeurIPS Best Paper award. Their later research includes language-model training-data extraction and the WMDP benchmark for malicious-use risks.

Now at
RunSybil
Works
11

Publication numbers from OpenAlex, as of 15 October 2026. Citations and h-index are hidden because this OpenAlex record may be incomplete or mix in other authors.

Timeline

  1. 2019

    Harvard and AI Village

    In 2019, Herbert-Voss is a Harvard computer-science PhD student and Berkman Klein Center fellow, studying adversarial machine learning, computer security, and mathematical optimization. They complete a double major in computer science and applied mathematics at the University of Utah before Harvard. They coauthor the 2019 report Release Strategies and the Social Impacts of Language Models and co-found DEF CON's AI Village.

  2. 2020

    GPT-3 and few-shot learning

    At OpenAI, Herbert-Voss coauthors Language Models are Few-Shot Learners, which trains GPT-3 as a 175-billion-parameter model and tests its few-shot performance. NeurIPS names the work among its 2020 Best Paper award recipients.

  3. 2021

    Extracting training data

    At USENIX Security in 2021, they coauthor a study showing that queries to GPT-2 can recover individual examples from its training data. The team extracts hundreds of text sequences, including personal information, and discusses safeguards.

  4. 2024

    The WMDP benchmark

    In 2024, Herbert-Voss coauthors The WMDP Benchmark, a test of language-model knowledge that could aid biological, cyber, and chemical weapons work. The paper examines machine unlearning as a way to reduce malicious-use capability.

  5. As of October 2026

    CEO and cofounder at RunSybil

    As of October 2026, Herbert-Voss is RunSybil's CEO and cofounder. The company builds autonomous agents that find security vulnerabilities in software.

Sources and credits

Photo: We have no published photo of Ariel Herbert-Voss yet. Send us a free-licensed photo.

  1. The right way to do AI in security | TechCrunch techcrunch.com. Checked 15 October 2026.
  2. About – Ariel Herbert-Voss – lab notebook adversariel.ai. Checked 15 October 2026.
  3. Berkman Klein Center Announces 2019-2020 Community | Berkman Klein Center cyber.harvard.edu. Checked 15 October 2026.
  4. [1908.09203] Release Strategies and the Social Impacts of Language Models arxiv.org. Checked 15 October 2026.
  5. Language Models are Few-Shot Learners proceedings.neurips.cc. Checked 15 October 2026.
  6. Announcing the NeurIPS 2020 award recipients – NeurIPS Blog blog.neurips.cc. Checked 15 October 2026.
  7. Extracting Training Data from Large Language Models | USENIX usenix.org. Checked 15 October 2026.
  8. [2403.03218] The WMDP Benchmark: Measuring and Reducing Malicious Use With Unlearning arxiv.org. Checked 15 October 2026.
  9. What Does It Take to Automate Hacker Intuition? | RunSybil Blog runsybil.com. Checked 15 October 2026.