Researcher

Anish Athalye

Anish Athalye is an AI researcher whose work spans adversarial machine learning, formal verification, and evaluation of AI agents. He earned MIT S.B. and S.M. degrees in 2017 and developed Porcupine, a Go linearizability checker; his 2018 paper on obfuscated gradients won ICML’s Best Paper Award. During his MIT PhD, he co-created Missing Semester, then co-founded Cleanlab and served as its CTO by 2023. His 2024 MIT thesis developed formal methods for checking secure hardware, and his 2026 work studied agents on investment-banking workflows and reinforcement learning with noisy rewards. As of October 2026, he works on ML research at Cursor.

Timeline

  1. 2017

    MIT degrees and Porcupine

    In 2017, he earns MIT S.B. and S.M. degrees. He develops Porcupine, a Go tool that checks concurrent-system histories for linearizability.

  2. 2018

    Obfuscated gradients at ICML

    With Nicholas Carlini and David Wagner, he shows in 2018 how obfuscated gradients can make adversarial-example defenses appear secure; their attacks circumvent six defenses completely and one partially. The paper wins ICML’s 2018 Best Paper Award.

  3. 2020

    The Missing Semester

    He co-creates MIT’s Missing Semester with Jon Gjengset and Jose Javier Gonzalez Ortiz, a course on programmer tools. Its 2020 lectures cover the shell, data wrangling, editors, and version control.

  4. By 2023

    Co-founding Cleanlab

    By 2023, he co-founds Cleanlab, a startup aimed at fixing data problems in AI models, and serves as its CTO.

  5. 2024

    Formal verification for secure hardware

    In 2024, his MIT CSAIL thesis develops information-preserving refinement and the Parfait framework to verify hardware security modules against leakage. He conducts this work in the Parallel and Distributed Operating Systems group, advised by Frans Kaashoek and Nickolai Zeldovich.

  6. January 2026

    Cleanlab joins Handshake

    In January 2026, Handshake acquires Cleanlab, and Athalye is set to join its AI Research & Strategy group with six researchers and engineers from frontier labs.

  7. April 2026

    Agentic benchmarks and noisy rewards

    In April 2026, he coauthors BankerToolBench, an open-source benchmark of end-to-end investment-banking workflows performed by junior bankers. With Andreas Plesner and Francisco Guzmán, he also studies reinforcement learning with verifiable rewards and finds validation reward within one percentage point of the clean baseline in tested Qwen3 8B trials with group-rollout noise through 20%.

  8. As of October 2026

    ML research at Cursor

    As of October 2026, he works on ML research at Cursor.

Career

  1. NowAnysphereML researcher
  2. CleanlabCo-founder, CTO

Education

  1. 2017Massachusetts Institute of TechnologyS.B. and S.M.

Sources and credits

Photo: We have no published photo of Anish Athalye yet. Send us a free-licensed photo.

  1. Clip Forbes: Cleanlab raises $25 million to solve AI models’ data mess (Oct 10) news.mit.edu. Checked 20 October 2026.
  2. GitHub - anishathalye/porcupine: A fast linearizability checker written in Go github.com. Checked 20 October 2026.
  3. Obfuscated Gradients Give a False Sense of Security: Circumventing Defenses to Adversarial Examples arxiv.org. Checked 20 October 2026.
  4. ICML 2018 Awards icml.cc. Checked 20 October 2026.
  5. 2020 Lectures · The Missing Semester of Your CS Education cgvr.informatik.uni-bremen.de. Checked 20 October 2026.
  6. Anish Athalye anish.io. Checked 20 October 2026.
  7. MIT CSAIL Parallel and Distributed Operating Systems Group: Publications vm-web.pdos.csail.mit.edu. Checked 20 October 2026.
  8. Handshake acquires Cleanlab joinhandshake.com. Checked 20 October 2026.
  9. BankerToolBench: Evaluating AI Agents in End-to-End Investment Banking Workflows arxiv.org. Checked 20 October 2026.
  10. An Imperfect Verifier is Good Enough: Learning with Noisy Rewards arxiv.org. Checked 20 October 2026.