Henry Papadatos

Hi, I’m Henry,

As a technical AI risk management expert, my focus lies in addressing the growing gap between rapid AI advancement and our ability to manage its associated risks.

I am the Executive Director of SaferAI where I work on both AI governance and technical solutions to improve the risk management for frontier AI systems.

Appointments

Member, EU AI Act Advisory Forum

Member, OECD Expert Group on AI Risk & Accountability

Expert, ISO/IEC JTC 1/SC 42 and AFNOR/CN IA, AI standards

Governance

I am contributing to various initiatives in the governance space. I was in an expert working group for the EU AI Act’s General-Purpose AI Code of Practice, focused on the risk taxonomy and on risk identification and assessment. My team took part in all four working groups.

I also took part in the OECD task force in charge of drafting the G7 Hiroshima AI Process reporting framework. Just before the Paris AI action summit, I advocated for international AI safety standards through an op-ed published in TIME.

Technical research

My team and I are building Europe’s capacity to evaluate frontier AI, working directly with model developers to evaluate their systems.

We also model AI risks across cyber, CBRN, and loss of control, with several AI safety institutes and the European Commission. Our work connects empirical measures of AI capabilities to estimates of real-world harm.

We built an AI risk management tracker for AI developers, featured in TIME and Euractiv, then a risk management framework that brings proven practices from other industries into AI.

Prior to joining SaferAI, I conducted technical research on large language models alignment at the Center for Human-Compatible AI at UC Berkeley.

Selected publications

Report

GLM-5.2 Risk Evaluation Report

Chinmayi Dixit^*, Jacob Davies^*, Jasmine Li, Ben Snodin, Jack Kengott, and 1 more author

2026

PDF

Paper

Lessons from External Review of DeepMind’s Scheming Inability Safety Case

Stephen Barrett, Francisco Javier Campos Zabala, Sean P. Fillingham, Umair Siddique, James Walpole, and 2 more authors

TAIGR Workshop, ICML, 2026

Best Paper arXiv Bib

Best Paper Award and selected for an oral presentation at the TAIGR Workshop, ICML 2026.

@article{barrett2026lessonsexternalreviewdeepminds,
  title = {Lessons from External Review of DeepMind's Scheming Inability Safety Case},
  author = {Barrett, Stephen and Zabala, Francisco Javier Campos and Fillingham, Sean P. and Siddique, Umair and Walpole, James and Bloomfield, Robin and Papadatos, Henry},
  year = {2026},
  journal = {TAIGR Workshop, ICML},
  eprint = {2604.21964},
  archiveprefix = {arXiv},
  primaryclass = {cs.CY},
  url = {https://arxiv.org/abs/2604.21964},
}

Paper

Toward Quantitative Modeling of Cybersecurity Risks Due to AI Misuse

Steve Barrett, Malcolm Murray, Otter Quarks, Matthew Smith, Jakub Kryś, and 15 more authors

2025

arXiv Bib

@article{barrett2025quantitativemodelingcybersecurityrisks,
  title = {Toward Quantitative Modeling of Cybersecurity Risks Due to AI Misuse},
  author = {Barrett, Steve and Murray, Malcolm and Quarks, Otter and Smith, Matthew and Kryś, Jakub and Campos, Siméon and Boria, Alejandro Tlaie and Touzet, Chloé and Hayrapet, Sevan and Heiding, Fred and Nevo, Omer and Swanda, Adam and Aguirre, Jair and Gershovich, Asher Brass and Clay, Eric and Fetterman, Ryan and Fritz, Mario and Juarez, Marc and Mavroudis, Vasilios and Papadatos, Henry},
  year = {2025},
  eprint = {2512.08864},
  archiveprefix = {arXiv},
  primaryclass = {cs.CY},
  url = {https://arxiv.org/abs/2512.08864},
}

Paper

A Frontier AI Risk Management Framework: Bridging the Gap Between Current AI Practices and Established Risk Management

Simeon Campos^*, Henry Papadatos^*, Fabien Roger, Chloé Touzet, Otter Quarks, and 1 more author

Conference on frontier AI safety frameworks (2024), 2025

arXiv Bib

@article{campos2025afrontierairisk,
  title = {A Frontier AI Risk Management Framework: Bridging the Gap Between Current AI Practices and Established Risk Management},
  author = {Campos, Simeon and Papadatos, Henry and Roger, Fabien and Touzet, Chloé and Quarks, Otter and Murray, Malcolm},
  year = {2025},
  eprint = {2502.06656},
  archiveprefix = {arXiv},
  primaryclass = {cs.AI},
  url = {https://arxiv.org/abs/2502.06656},
  journal = {Conference on frontier AI safety frameworks (2024)},
}

Paper

Linear Probe Penalties Reduce LLM Sycophancy

Henry Papadatos, and Rachel Freedman

SoLaR workshop, NeurIPS, 2024

arXiv Bib

@article{papadatos2024linearprobepenaltiesreduce,
  title = {Linear Probe Penalties Reduce LLM Sycophancy},
  author = {Papadatos, Henry and Freedman, Rachel},
  year = {2024},
  eprint = {2412.00967},
  archiveprefix = {arXiv},
  primaryclass = {cs.AI},
  journal = {SoLaR workshop, NeurIPS},
}

All publications →