PeopleofSuperintelligence

Researcher

Daniel Paleka

Daniel Paleka is an AI security researcher whose work spans neural-network theory, model safety and privacy. He won silver at the International Mathematical Olympiad in 2016, authored programming-contest problems, and studied deep ReLU-network injectivity in a 2021 master’s thesis. His 2022 paper examined the Stable Diffusion safety filter, and his 2023 paper studied web-scale data poisoning; his model-stealing paper won an ICML 2024 Best Paper award, and his 2025 work covered forecasting evaluation and safety tooling. As of October 2026, he is a PhD student at ETH Zurich’s Secure and Private AI Lab, where his research includes LLM deanonymization and distillation-defense threat models.

Timeline

  1. 2016

    Silver at the Hong Kong IMO

    In 2016, he wins silver at the International Mathematical Olympiad in Hong Kong after earlier silver medals in Cape Town and Thailand. The Zadar student prepares to study mathematics at the University of Zagreb’s Faculty of Science in autumn 2016.

  2. 2016–2018

    Authoring contest problems

    He authors or coauthors problems for COCI rounds in 2016/2017 and 2017/2018, as well as Microsoft Bubble Cup 2018.

  3. November 2021

    A thesis on ReLU networks

    By November 2021, his master’s thesis reports the first theorem on injectivity of deep ReLU neural networks at initialization. The work builds on joint research with Afonso Bandeira, Charles Clum and Dustin Mixon.

  4. 2022

    A PhD and a safety-filter study

    In September 2022, he begins a PhD advised by Florian Tramèr on security and failure modes of artificial intelligence. He coauthors Red-Teaming the Stable Diffusion Safety Filter, which shows that the filter can be bypassed and ignores violence and gore.

  5. 2023

    Poisoning web-scale datasets

    In 2023, he coauthors the preprint Poisoning Web-Scale Training Datasets is Practical, which presents two attacks that could poison ten popular datasets.

  6. 2024

    A Best Paper on model stealing

    In 2024, he coauthors Stealing Part of a Production Language Model, which recovers a transformer’s embedding projection layer through API access and extracts the full Ada and Babbage projection matrices. ICML gives it a Best Paper award.

  7. 2025

    Forecasting and safety tooling

    In 2025, he coauthors Pitfalls in Evaluating Language Model Forecasters, which identifies temporal leakage and difficulty in extrapolating benchmark performance to real-world forecasting. He also co-creates safety-research/safety-tooling v1.0.0, released in May 2025.

  8. As of October 2026

    PhD student at SPY Lab

    As of October 2026, he is a PhD student in ETH Zurich’s Secure and Private AI Lab (SPY Lab). In 2026, he coauthors Large-scale online deanonymization with LLMs, a study that re-identifies people from pseudonymous profiles and conversations, and What Does It Mean to Break a Distillation Defense?, which proposes a threat model for API-accessible LLMs.

  1. The median h-index is 48 at universities and 14 at frontier AI labsIncludes ETH Zurich. The median h-index is 48 at universities and 14 at frontier AI labs. We compare publishing careers, recent citations, dual appointments and job moves across 2,420 university researchers and 308 frontier-lab researchers.
  2. The Massachusetts Institute of Technology trained or employed 44 of the 187 AI major-prize winners, more than any other universityIncludes ETH Zurich. The Massachusetts Institute of Technology trained or employed 44 of the 187 AI major-prize winners, more than any other university. We compare doctorate and 5+ year employment links, career timing and current workplaces for 187 major-prize and 1,111 best-paper winners.
  3. The company share of AI researchers’ top papers was 31% in 2020–2024 and 8% in 1970–1974Includes ETH Zurich. We estimate where the people shaping AI worked when their most-cited papers appeared: the company share was 31% in 2020–2024 and 8% in 1970–1974, and Google led all organizations in 2020–2024.
  4. Universities had a net outflow of 2,068 moves in AI careersIncludes ETH Zurich. Our estimates of career flows and median h-index across universities and companies in AI, including the role of joint appointments and earlier academic jobs.

Colleagues

People named in this biography, and people with dated positions at the same organization in the same years.

Sources and credits

Photo: We have no published photo of Daniel Paleka yet. Send us a free-licensed photo.

  1. Srebrni Daniel Paleka: 'Svi matematički zadaci u Hong Kongu bili su teški. I neka su.' srednja.hr. Checked 22 October 2026.
  2. Competitive Programming danielpaleka.com. Checked 22 October 2026.
  3. Interesting problems in DL theory danielpaleka.com. Checked 22 October 2026.
  4. Daniel Paleka danielpaleka.com. Checked 22 October 2026.
  5. Red-Teaming the Stable Diffusion Safety Filter arxiv.org. Checked 22 October 2026.
  6. Poisoning Web-Scale Training Datasets is Practical arxiv.org. Checked 22 October 2026.
  7. Stealing Part of a Production Language Model arxiv.org. Checked 22 October 2026.
  8. ICML 2024 Awards icml.cc. Checked 22 October 2026.
  9. Pitfalls in Evaluating Language Model Forecasters arxiv.org. Checked 22 October 2026.
  10. [CODE] safety-research/safety-tooling: v1.0.0 research.ed.ac.uk. Checked 22 October 2026.
  11. SPY Lab spylab.ai. Checked 22 October 2026.
  12. Publications | SPY Lab spylab.ai. Checked 22 October 2026.
  13. Large-scale online deanonymization with LLMs arxiv.org. Checked 22 October 2026.
  14. What Does It Mean to Break a Distillation Defense? arxiv.org. Checked 22 October 2026.