I had the pleasure to give a talk, titled “Trustworthy Machine Learning in the era of Large Language Models”, at CENIA (Santiago, Chile) about our work at Parameter Lab.
| CARVIEW |
Select Language
HTTP/2 200
cache-control: max-age=600
content-type: text/html; charset=utf-8
etag: "d8598429e028aedba2753a9846e516039a9eb69d450af140d21e5bf3edf41aa4"
expires: Tue, 30 Dec 2025 10:01:02 CET
last-modified: Sun, 28 Dec 2025 23:07:34 GMT
permissions-policy: browsing-topics=()
strict-transport-security: max-age=63072000; preload
vary: Origin
x-request-id: 01KDQ787TH30VW49HT6A2W5CG2
content-length: 53874
date: Tue, 30 Dec 2025 08:51:02 GMT
Martin Gubri's Website
Biography
I am the research lead at Parameter Lab in Tübingen, Germany. My work focuses on auditing the hidden risks of black-box AI systems to make them more safe, transparent, and accountable. Through systematic external testing of large language models (LLMs) and computer vision models, I develop methods to independently assess risks and ensure compliance without requiring access to model internals.
I hold a PhD in computer science from the University of Luxembourg, where my thesis focused on adversarial machine learning. I am a data scientist who graduated from Ensae ParisTech and a statistician who graduated from Toulouse School of Economics.
I am considering opportunities in both academia and industry. If my profile matches what you are looking for, I would be glad to hear from you.
You can view my CV here.
Interests
- Trustworthy AI
- Machine Learning Auditing
- Large Language Model
- Adversarial Machine Learning
- Privacy and Security
Education
-
PhD in Computer Science, 2023
University of Luxembourg (Luxembourg)
-
Specialized Master's in Data Science, 2015
Ensae ParisTech (France)
-
Dual Master's Degrees in Statistics and Econometrics, 2014
Toulouse School of Economics and Paul Sabatier University (France)
-
Bachelor's in Economics and Mathematics, 2012
Toulouse School of Economics (France)
News
December 2025
- I have been appointed Area Chair for UAI 2026.
- I will attend NeurIPS 2025 in San Diego to present C-SEO Bench
November 2025
- I will give my talk titled “Revealing the Invisible: Auditing the Hidden Risks of Black-Box LLMs” at the University of Luxembourg (Luxembourg) on November 20th at 3pm
October 2025
- I have been appointed Area Chair for ARR ACL.
- I had the pleasure of giving a talk titled “Revealing the Invisible: Auditing the Hidden Risks of Black-Box LLMs” at the University of Mannheim (Germany) and the University of Trento (Italy)
- Our new Dr.LLM paper was selected as one of the top AI papers of the week by DAIR.AI!
September 2025
- Two papers that I had the pleasure of supervising were accepted: C-SEO Bench at NeurIPS 2025 D&B and Leaky Thoughts at EMNLP 2025
June 2025
- Our new Leaky Thoughts paper was selected as one of the top AI papers of the week by DAIR.AI and The AI Timeline!
Selected Publications
TRAP: Targeted Random Adversarial Prompt Honeypot for Black-Box Identification
The paper introduces Targeted Random Adversarial Prompt (TRAP), an LLM fingerprinting method, to identify a specific LLM in third-party applications.
Martin Gubri, Dennis Ulmer, Hwaran Lee, Sangdoo Yun, Seong Joon Oh
ACL 2024 (findings)
Publications
- Is Multilingual LLM Watermarking Truly Multilingual? A Simple Back-Translation Solution
- DISCO: Diversifying Sample Condensation for Accelerating Model Evaluation
- Dr.LLM: Dynamic Layer Routing in LLMs
- Leaky Thoughts: Large Reasoning Models Are Not Private Thinkers
- C-SEO Bench: Does Conversational SEO Work?
- Social Science Is Necessary for Operationalizing Socially Responsible Foundation Models
- Testing Uniform Random Samplers: Methods, Datasets and Protocols
- Scaling Up Membership Inference: When and How Attacks Succeed on Large Language Models
- Calibrating Large Language Models Using Their Generations Only
- TRAP: Targeted Random Adversarial Prompt Honeypot for Black-Box Identification
- ProPILE: Probing Privacy Leakage in Large Language Models
- What Matters in Model Training to Transfer Adversarial Examples
- Going Further: Flatness at the Rescue of Early Stopping for Adversarial Example Transferability
- LGV: Boosting Adversarial Example Transferability from Large Geometric Vicinity
- Efficient and Transferable Adversarial Examples from Bayesian Neural Networks
- Influence-driven data poisoning in graph-based semi-supervised classifiers
- Search-Based Adversarial Testing and Improvement of Constrained Credit Scoring Systems
- Adversarial perturbation intensity strategy achieving chosen intra-technique transferability level for logistic regression
Recent Posts
Cyberwal 2022 Winter School on Machine Learning Security in the Real World
Content of the cyberwal 2022 workshop on machine learning security in the real world.
Development
FLOSS Contributions
Technical Supervisor
- MASEval. Supervised the development of MASEval, an open-source LLM-based multi-agent evaluation framework.
Significant Contributions
- Adversarial Robustness Toolbox (ART)
- Torchattacks
- Implementation of state-of-the-art predictors for Spatial Regression Models in the spdep R package
- Integration of sentences in French into Common Voice (epub parsing and cleaning)
Minor Contributions
Teaching
2023 – 2025
Master: Erasmus Mundus Joint Master in Cybersecurity
The Erasmus Mundus Joint Master in cybersecurity (CYBERUS) at the University of Luxembourg. One session about adversarial examples against LLMs per year (2023, 2024, 2025).
Before Class: Preparation (15 mins)
Read these short resources before our session:
- GCG paper abstract (2 mins)
- Watch: What Is a Prompt Injection Attack? (10 mins)
- LLM alignment defintion (3 mins)
Reflection question: Why LLM developers might want to prevent certain outputs from their LLMs?
Class Content
- Slides.
- Notebook showcasing jailbreaking: Google Colab, ipynb file.
2022
Master: Advanced topics in Applied Machine Learning
2nd year of Master in Computer Science. Two lectures, design and correction of the project, planning.
- Recalls of machine learning, machine learning frameworks, first part project
- Overview of adversarial machine learning, model calibration
Project: Creation and evaluation of fill-in-the-blank notebooks (part 1 on preprocessing and part 3 on adversarial examples)
Master: Introduction to Machine Learning
2nd year of Master in Space Science. Six sessions, including four based on the Machine Learning Refined book and one based on the Applied Machine Learning course of Andreas C. Müller. Summary slides.
- Recalls of linear algebra: “Machine Learning Refined” book, “Essence of linear algebra” videos serie of 3blue1brown, some formal definitions from the “Mathematics for Machine Learning” book.
- Zero-order optimization: Chapter 2, Appendix B.
- First-order optimization: Chapter 3, “Gradient descent, how neural networks learn” video from 3Blue1Brown.
- Linear regression and linear classification: Chapters 5 and 6
- ML project lifecycle: Data preparation, feature engineering, overfitting & underfitting, model evaluation. Slides.
- Neural Networks (slides), Keras & Convolutional Neural Nets (slides) and Advanced Neural Networks (slides).
- Written examination
2021
Master: Introduction to Machine Learning
2nd year of Master in Space Science. Two introductory lectures on Machine Learning. Slides.
2020
Bachelor: Software engineering 2
3rd year of Bachelor in Computer Science. Four introductory lectures on Machine Learning Engineering. Course given online during lockdown. Quizzes on Moodle. Videos, Slides
- Introduction to Machine Learning: Useful Definitions, Types of Tasks in Machine Learning
- Introduction to Machine Learning: Recalls of Statistics, Model’s elements, Elements of Statistical Learning Theory
- Machine Learning Project Lifecycle: When to (not) use Machine Learning, Goal Definition, Data Collection & Preparation
- Machine Learning Project Lifecycle: Feature Engineering, Choosing and Training a model, Model Evaluation, Feedback loop
Academic Services
Meta-Reviewer
- Area Chair for UAI 2026
- Area Chair for ARR ACL (Oct. 2025, Jan. 2026)
Reviewer
I served as a reviewer for the following venues.
- ACL ARR (May 2025)
- UAI (2024, 2023)
- International Journal of Computer Vision
- IEEE Transactions on Neural Networks and Learning Systems
- IEEE Transactions on Pattern Analysis and Machine Intelligence
- SiMLA Workshop (2023)
Miscellaneous
I organized and animated the weekly Machine Learning Reading Group at the SerVal group (University of Luxembourg) from February 2021 to August 2023.
White-Hat
Contributions to FLOSS Security
Vulnerabilities Discovered
| CVE | Software | Type | Description/Impact | Links |
|---|---|---|---|---|
| CVE-2017-6877 | Lutim | Stored XSS | Exposed all images uploaded by the user and their encryption keys | issue |
| CVE-2017-10975 | Lutim | Stored XSS | Idem. Hard to exploit in pratice | issue |
| CVE-2017-1000051 | CryptPad | Stored XSS | Exposed encryption keys of user data | blog post |
| CVE-2017-11594 | Loomio | Stored XSS | Markdown not sanitized. Allows to cast users’ votes using their identity | commit, demo |
| Loomio | Stored XSS | No restrictions to attached files (when served locally). Allows to cast users’ votes using their identity | demo | |
| CVE-2017-1000039 | Framadate | Formula Injection | issue, MR | |
| CVE-2017-11593 | Markdown Preview Plus Chrome’s Extension | Stored XSS | Led its users vulnerable to XSS in a ton of websites, by converting text, markdown and rst files to HTML without sanitization | issue |
| CVE-2016-6127 | Request Tracker (RT) | Stored XSS | Exposed all users data to remote attacker. Concurrent finding. | |
| TeleR | RCE | 3 Arbitrary Code Executions on their server | ||
| Turtl | Stored XSS | 3 XSS exposing encrypted data (incl. passwords) | ||
| NCrypt | Stored XSS | issue | ||
| Sympa | Stored XSS | No public record | ||
| GNU Social | Stored XSS | No public record | ||
| Shaarli | Stored XSS | Markdown plugin | MR | |
| Drupal/webform | Stored XSS | Exposed user and admin data when visiting an uploaded file. Concurrent finding. | issue | |
| Framaforms | Improper Access Control | Exposed URL of all users forms | No public record | |
| Framaforms | Stored XSS | Exposed responses of user forms. Too permissive formats allowed to untrusted users | issue | |
| Framaforms | Stored XSS | issue | ||
| Framaslides | Stored XSS | Markdown not sanitized | commit | |
| Framaslides | Stored XSS | Escape markdown link sanitization (marked lib not updated) | issue | |
| Framaslides | Stored XSS | issue | ||
| Framemo & Sandstorm’s Scrumblr | Stored XSS | Markdown not sanitized | issue, PR | |
| Framemo & Sandstorm’s Scrumblr | Formula Injection | issue, MR | ||
| Framaboard | Stored XSS | Affected Framaboard, but default Kanboard not vulnerable | commit | |
| Wallabag 2 & Graby | Stored XSS | PR | ||
| Kresus | Stored Self-XSS | Possible to leverage it by importing a malicious JSON | issue | |
| Dolomon | Stored (Self)-XSS | Multiple XSS. Some can be leveraged using a CSRF issue | issue | |
| Dolomon | Improper Access Control | Gave access to the URLs saved by all users | issue | |
| Dolomon | Formula Injection | issue | ||
| Mastodon | Mime-sniffing | Low (only Internet Explorer affected) | No public record | |
| share-on-diaspora Wordpress Plugin | Reflected Client XSS | Fixed, but not discovered. | PR |