Chuan Guo

Timeline

1. By 2015

   Mathematics and CS at Waterloo

   Guo earns bachelor's and master's degrees in Mathematics and Computer Science at the University of Waterloo, advised by Douglas R. Stinson. The 2015 master's thesis is titled Fingerprinting Codes and Related Combinatorial Structures.

2. 2017

   Calibrating modern neural networks

   At Cornell, where Kilian Weinberger and Karthik Sridharan co-advise their PhD, Guo co-writes On Calibration of Modern Neural Networks with Geoff Pleiss, Yu Sun and Weinberger for ICML 2017. The paper finds that modern neural networks are poorly calibrated and that temperature scaling, a single-parameter variant of Platt scaling, is surprisingly effective at calibrating predictions.

3. 2018–2019

   Input transformations and black-box attacks

   In 2018, Guo co-writes Countering Adversarial Images using Input Transformations with Mayank Rana, Moustapha Cisse and Laurens van der Maaten; the best defense in the study eliminates 60% of strong gray-box and 90% of strong black-box attacks. In 2019, Simple Black-box Adversarial Attacks, written with Jacob R. Gardner, Yurong You, Andrew Gordon Wilson and Kilian Q. Weinberger, proposes a simple method for building adversarial images in the black-box setting.

4. 2020–2025

   Research scientist at Meta FAIR

   Guo completes a Cornell PhD in 2020 with the thesis Threats and Countermeasures in Machine Learning Applications, then works as a research scientist on Meta's Fundamental AI Research (FAIR) team until 2025. The 2020 ICML paper Certified Data Removal from Machine Learning Models, written with Tom Goldstein, Awni Hannun and Laurens van der Maaten, defines certified removal and develops a removal mechanism for linear classifiers.

5. 2021

   A UAI best paper

   Measuring Data Leakage in Machine-Learning Models with Fisher Information, written with Awni Hannun and Laurens van der Maaten, wins a best paper award at UAI 2021. The paper proposes quantifying leakage with the Fisher information of the model about the data.

6. 2025

   SecAlign and Meta-SecAlign

   SecAlign, a prompt-injection defense built on preference optimization, appears at the ACM Conference on Computer and Communications Security in 2025, with Sizhe Chen, Arman Zharmagambetov, Saeed Mahloujifar, Kamalika Chaudhuri and David Wagner as co-authors. In July 2025, the Meta-SecAlign paper, written with Chen, Zharmagambetov and Wagner, introduces a utility-preserving defense and links code plus 8B and 70B models.

7. 2026

   ICML honorable mention on memorization

   How much can language models memorize?, written with John Morris, Chawin Sitawarin, Narine Kokhlikyan, Edward Suh, Alexander Rush, Kamalika Chaudhuri and Saeed Mahloujifar, receives an Outstanding Paper Honorable Mention at ICML 2026. It estimates that GPT-style models have a capacity of approximately 3.6 bits per parameter.

8. As of October 2026

   Researcher at OpenAI

   Guo works as a researcher at OpenAI on AI security and privacy, with adversarial robustness, AI agent security and privacy-preserving machine learning among the active topics. In March 2026, the IH-Challenge paper, with Guo as first author, introduces a reinforcement learning training dataset for instruction hierarchy; fine-tuning GPT-5-Mini on it raises instruction-hierarchy robustness from 84.1% to 94.1% on average across 16 benchmarks.