Samuel Šimko

I am an ELLIS PhD student working on AI safety, advised by Prof. Bernhard Schölkopf and co-advised by Prof. Zhijing Jin. Before that, I completed my MSc in Computer Science at ETH Zurich.

My research is about making language models hold up under adversarial pressure: tamper-resistant safeguards, defenses that survive fine-tuning, and robustness to jailbreaking and prompt injection attacks. More broadly, I am interested in adversarial defenses, causality, representation learning, and misalignment detection. My work has been featured at ICML, COLM, EMNLP and IASEAI. In the past, I have also worked on projects in computational cosmology and AI for healthcare.

Outside of research, I am an avid speedcuber. I have also participated and won prizes in several GraySwan Arena manual jailbreaking contests.

You can find my latest CV here

News

Sep 01, 2026	Excited to be starting my ELLIS PhD, advised by Prof. Bernhard Schölkopf and Prof. Zhijing Jin!
Jul 26, 2026	Happy to have taken part in the MARS 5.0 research sprint in Oxford, UK 🇬🇧.
Jul 08, 2026	Excited to be at ICML 2026 in Seoul, South Korea 🇰🇷, presenting Training with Honeypots!
Jun 07, 2026	Happy to have attended the CHAI 2026 workshop in Asilomar, California 🇺🇸.
Feb 26, 2026	Happy to have attended IASEAI’26 at UNESCO House in Paris, France 🇫🇷.

Selected Publications

Improved Robustness Against Indirect Prompt Injection Can Be Built into LLMs

M. Özdinçer, Samuel Simko, Bernhard Schölkopf, and Zhijing Jin

In Conference on Language Modeling (COLM), 2026

COLM 2026 Bib

Accepted to COLM 2026

@inproceedings{ozdincer2026injection,
  title = {Improved Robustness Against Indirect Prompt Injection Can Be Built into LLMs},
  author = {{\"O}zdin{\c c}er, M. and Simko, Samuel and Sch{\"o}lkopf, Bernhard and Jin, Zhijing},
  booktitle = {Conference on Language Modeling (COLM)},
  year = {2026},
}

Training with Honeypots: Reshaping How LLMs Fail Under Adversarial Attacks

Samuel Simko, Punya Syon Pandey, Zhijing Jin, and Bernhard Schölkopf

In International Conference on Machine Learning (ICML), 2026

ICML 2026 Bib HTML Code

Accepted to ICML 2026

@inproceedings{simko2026honeypots,
  title = {Training with Honeypots: Reshaping How LLMs Fail Under Adversarial Attacks},
  author = {Simko, Samuel and Pandey, Punya Syon and Jin, Zhijing and Sch{\"o}lkopf, Bernhard},
  booktitle = {International Conference on Machine Learning (ICML)},
  year = {2026},
}

Position: Safe Models Do Not Guarantee Safe Societies: The Case for Sociopolitical Risk

David Guzman Piedrahita, Changling Li, Dave Banerjee, Terry Jingchen Zhang, and 7 more authors

In International Conference on Machine Learning (ICML), Position Paper Track, 2026

ICML 2026 Spotlight Bib HTML

Accepted as a spotlight to the ICML 2026 Position Paper Track

@inproceedings{piedrahita2026position,
  title = {Position: Safe Models Do Not Guarantee Safe Societies: The Case for Sociopolitical Risk},
  author = {Guzman Piedrahita, David and Li, Changling and Banerjee, Dave and Zhang, Terry Jingchen and Blin, Kevin and Simko, Samuel and Pandey, Punya Syon and Strauss, Irene and Mihalcea, Rada and Sch{\"o}lkopf, Bernhard and Jin, Zhijing},
  booktitle = {International Conference on Machine Learning (ICML), Position Paper Track},
  year = {2026},
}

Improving Large Language Model Safety with Contrastive Representation Learning

Samuel Simko, Mrinmaya Sachan, Bernhard Schölkopf, and Zhijing Jin

arXiv preprint arXiv:2506.11938, 2025

EMNLP 2025 arXiv Bib

Accepted as a poster to EMNLP Main 2025

@article{simko2025improving,
  title = {Improving Large Language Model Safety with Contrastive Representation Learning},
  author = {Simko, Samuel and Sachan, Mrinmaya and Sch{\"o}lkopf, Bernhard and Jin, Zhijing},
  journal = {arXiv preprint arXiv:2506.11938},
  year = {2025},
}