Aleksander Mądry
Last reviewed
Jun 7, 2026
Sources
11 citations
Review status
Source-backed
Revision
v1 · 1,541 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Jun 7, 2026
Sources
11 citations
Review status
Source-backed
Revision
v1 · 1,541 words
Add missing citations, update stale details, or suggest a clearer explanation.
Aleksander Mądry (born in Wrocław, Poland) is a Polish American computer scientist, the Cadence Design Systems Professor of Computing in the Department of Electrical Engineering and Computer Science at the Massachusetts Institute of Technology (MIT) and a member of the MIT Computer Science and Artificial Intelligence Laboratory (CSAIL). He is widely known for foundational work in adversarial machine learning and reliable machine learning, and for earlier breakthroughs in algorithmic graph theory and optimization. He directs the MIT Center for Deployable Machine Learning and leads the research group commonly called the Madry Lab. Since 2023 he has been on leave from MIT to work at OpenAI, where he served as the company's first Head of Preparedness before moving to a research role focused on reasoning. [1][2]
Mądry grew up in Wrocław and studied physics and computer science at the University of Wrocław. By his own account he first took up programming in order to build video games, then was drawn instead toward theoretical computer science, algorithms, and optimization. [2]
He moved to the United States for graduate study at MIT, earning his PhD in computer science in 2011 under the supervision of Michel X. Goemans and Jonathan A. Kelner. His dissertation, "From Graphs to Matrices, and Back: New Techniques for Graph Algorithms," developed methods that bring continuous optimization and linear algebra to bear on classical combinatorial graph problems. The thesis received MIT's George M. Sprowls Award for the best doctoral thesis in computer science and an honorable mention for the ACM Doctoral Dissertation Award. [2][7]
After completing his doctorate, Mądry spent a year as a postdoctoral researcher at Microsoft Research New England, then joined the faculty of the Ecole Polytechnique Federale de Lausanne (EPFL) in Switzerland, where he remained until early 2015. A common misconception holds that his doctorate is from EPFL; in fact EPFL was his first faculty appointment, and his PhD is from MIT. [1]
In 2015 Mądry returned to MIT as a member of the EECS faculty, joining CSAIL and its Theory of Computation Group. He was later promoted to a tenured professorship, and in 2021 he was named the inaugural Cadence Design Systems Professor of Computing, a chair established to support faculty working in artificial intelligence, machine learning, and data analytics within the MIT Stephen A. Schwarzman College of Computing. [3]
At MIT he founded and directs the Center for Deployable Machine Learning, a research effort launched under the MIT Quest for Intelligence with the goal of making machine learning dependable enough for high-stakes settings such as autonomous vehicles and health care. He also serves as a faculty co-lead of the MIT AI Policy Forum, which connects technical research with policymaking. [1][2]
Mądry's early research lay in theoretical computer science, particularly fast algorithms for graph problems. He is best known in that area for a 2013 algorithm for the maximum flow problem that broke a long-standing running-time barrier: for unit-capacity networks he gave a method running in roughly O(m^{10/7}) time, improving on bounds that had stood for decades on sparse graphs. The work combined ideas from continuous optimization, electrical flows, and interior-point methods, and it helped set off a wave of progress on flow and related problems. For this and related contributions he received the 2018 Presburger Award, given by the European Association for Theoretical Computer Science to an outstanding young scientist. [6]
In the late 2010s Mądry shifted much of his attention to the reliability of deep learning. With collaborators he studied adversarial examples: inputs altered by small, often imperceptible perturbations that cause models to make confident mistakes. Their 2017 paper "Towards Deep Learning Models Resistant to Adversarial Attacks" recast the problem as one of robust optimization, framing robust training as a saddle-point, or min-max, problem. The paper popularized projected gradient descent (PGD) as a strong first-order attack and proposed PGD-based adversarial training, which remains one of the most widely used and durable methods for hardening neural networks. It has been cited many thousands of times and is a standard reference in adversarial machine learning. [4]
A second influential result, "Adversarial Examples Are Not Bugs, They Are Features" (NeurIPS 2019), argued that adversarial vulnerability is not merely a quirk of imperfect models but a consequence of "non-robust features": patterns in data that are genuinely predictive yet brittle and meaningless to humans. The paper reframed robustness as partly a question of which features a model should be allowed to use, and it influenced later work on interpretability and on the alignment between human and machine notions of meaning. [5]
Across these projects, Mądry has argued for doing machine learning "the right way," meaning building models whose behavior people can understand, predict, and trust before deploying them in the real world. This theme of dependable, deployable machine learning ties his academic work to his later industry role in AI safety. [2]
In 2023 Mądry took leave from MIT to join OpenAI. In October 2023 the company announced the formation of a new Preparedness team, with Mądry as its head. The team's remit was to track, evaluate, forecast, and protect against catastrophic risks from frontier AI models, spanning both misuse and emergent capabilities. [9][10]
Under Mądry the team authored OpenAI's Preparedness Framework, first released in beta in December 2023. The Preparedness Framework set out a process for measuring dangerous capabilities in frontier models and gating their development and deployment accordingly. The original version tracked four risk categories: cybersecurity; chemical, biological, radiological, and nuclear (CBRN) threats; persuasion; and model autonomy. Each was assigned a risk level of low, medium, high, or critical; models scoring "high" in any category could not be deployed, and a "critical" score was meant to halt further development until safeguards were in place. OpenAI later revised the framework in 2025. [9]
In July 2024 OpenAI reassigned Mądry from the Preparedness role to a research role focused on AI reasoning. The change drew press attention because it came shortly after U.S. senators had written to OpenAI about its safety practices, though the company described the move as a research reassignment rather than a demotion. Oversight of the preparedness work passed to Joaquin Quinonero Candela and Lilian Weng. [10]
As of 2026 Mądry remains on leave from MIT and continues to work at OpenAI on reasoning research. The Head of Preparedness position has since seen further turnover, and in late December 2025 OpenAI chief executive Sam Altman publicly announced a renewed search for someone to fill the role, citing the growing challenges posed by advancing models. [11]
Mądry has been an active voice in public debates over AI policy. On March 10, 2023, testifying before the U.S. House Subcommittee on Cybersecurity, Information Technology, and Government Innovation, he warned that government should not "abdicate" responsibility for AI's direction to private companies, cautioned that people tend to over-trust human-sounding systems, and said that society had reached "an inflection point" requiring careful public discussion. He has also submitted testimony in other governmental settings, including a written statement provided in his capacity as Head of Preparedness at OpenAI. [8]
| Year | Milestone |
|---|---|
| 2011 | PhD in computer science, MIT (advisors Goemans and Kelner); George M. Sprowls Award |
| 2011 to 2012 | Postdoctoral researcher, Microsoft Research New England |
| 2012 to 2015 | Faculty member, EPFL (Lausanne, Switzerland) |
| 2015 | Joins MIT EECS faculty and CSAIL |
| 2017 to 2018 | "Towards Deep Learning Models Resistant to Adversarial Attacks" (PGD adversarial training) |
| 2018 | Presburger Award, EATCS |
| 2019 | "Adversarial Examples Are Not Bugs, They Are Features," NeurIPS |
| 2021 | Named inaugural Cadence Design Systems Professor of Computing |
| 2023 | Joins OpenAI; named Head of Preparedness (team announced October 2023) |
| 2024 | Reassigned to an AI reasoning research role (July) |
| 2025 to 2026 | Continues reasoning research at OpenAI; remains on leave from MIT |