About

Hi, I’m Thomas! I’m currently an M2 mathematics student at Télécom SudParis and ENS Paris-Saclay, and soon, an AI researcher.

I began doing research in AI security two years ago. After taking a gap year, where I studied Large Language Model (LLM) security and LLM interpretability, I’m now turning my attention toward more fundamental AI. I also love graph theory and physics.

My favorite paper: Toy Models of Superposition.

You can find me on:

  • GitHub / Sckathach
  • LinkedIN / thomas-winninger
  • X / sckathach

You can contact me at: thomas [dot] winninger [at] telecom-sudparis [dot] eu

Download PDF resume: in English.

See my Research work.

Education

  • 2025 - 2026 - Master MVA, ENS Paris-Saclay
    Topology, optimal transport, reinforcement learning, training and deploying large-scale models, LLM, graph neural networks, learning for protein science, convex optimization.

  • 2022 - 2026 - Engineering Degree, Télécom SudParis
    Telecommunications, cyber security, cloud, information theory, probability, optimization, graph theory, graph neural networks, signal processing.

Experience

  • Sep 2025 - now - Teaching and research sprints - PIAF
    Teaching (interpretability, LLM training and fine-tuning) and organizing short research sprints (teams of ~5 people, lasting under four days).

  • Jul - Sep 2025 - Research internship in LLM security - NICT
    Research on the security and jailbreak interpretability of Large Reasoning Models (LRMs). I studied LRM robustness, adapted state-of-the-art black-box and white-box attack from LLMs, and started studying jailbreaks with interpretability methods on LRMs.

  • Mar - May 2025 - Research internship in AI explanability - INRIA
    Verified robust explanation for language models. I explored scaling Hybrid Constrained Zonotopes (HCZs) to language models using convex relaxation and optimization. However, the relaxation error proved too large for practical use.

  • Jul - Dec 2024 - Research internship in AI security - Thales
    Implementations and improvements of state-of-the-art attacks on LLMs. I improved state-of-the-art white-box adversarial attacks on LLMs and published the results on ArXiv.

  • 2022 - 2024 - Teaching and infrastructure - HackademINT
    Teaching (cloud and AI security), cloud management (Kubernetes), creation of challenges (AI & quantum physics), and organization of 404CTF 2023 & 2024 (largest cyber security competition in France).

Miscellaneous

  • Languages: Python, French, OCaml, English, Typst, TypeScript, Lua, Rust, C, Bash, Japanese (JLPT 4), Lean
  • Tools/ Frameworks: PyTorch, nnsight, Docker (Podman), Kubernetes, React, Qiskit, Archlinux
  • Other interests: Piano, guitar, teaching, reading, geopolitics, particle physics, sports, video game (playing & development), meditation
  • I completed the Alignment Research Engineer Accelerator (ARENA) and the AI Safety Fundamental (AISF) curriculums.