PeopleofSuperintelligence

Researcher

Hjalmar Wijk

As of October 2026, Hjalmar Wijk is chief scientist at METR, where he works on threat modeling to determine which AI capabilities may pose catastrophic risk. His earlier AI risk and strategy work was at the Future of Humanity Institute and the Center on Long-Term Risk, where he was a 2022 summer research fellow. He co-wrote a 2021 paper on ensuring the safety of deep reinforcement learning agents in Atari games. In November 2024 he helped write RE-Bench, which sets AI agents against data from 61 distinct human experts. In August 2026 he and Ajeya Cotra spent a total of six days on OpenAI's premises, and METR published their investigation of agent behavior in the Hugging Face hacking incident.

Now at
METR
Works
10
Citations
37
h-index
3

Publication numbers from OpenAlex, as of 15 October 2026.

Timeline

  1. 2021

    Shielding Atari games

    He co-writes Shielding Atari Games with Bounded Prescience alongside Mirco Giacobbe, Mohammadhosein Hasanbeig and Daniel Kroening. The paper presents a method for ensuring the safety of deep reinforcement learning agents in Atari games.

  2. 2022

    Summer research fellow at CLR

    He is one of the 2022 summer research fellows at the Center on Long-Term Risk (CLR). He has previously worked on AI risks and strategy at the Future of Humanity Institute and at CLR.

  3. December 2023

    Autonomous replication and adaptation

    He co-authors Evaluating Language-Model Agents on Realistic Autonomous Tasks, which explores the ability of language model agents to acquire resources, create copies of themselves, and adapt to novel challenges.

  4. November 2024

    The RE-Bench benchmark

    He co-writes RE-Bench, a benchmark of seven open-ended machine learning research engineering environments. It sets AI agents against data from 61 distinct human experts.

  5. March 2025

    A time-horizon metric

    He co-authors Measuring AI Ability to Complete Long Software Tasks, which proposes a 50%-task-completion time horizon as a new metric for AI models.

  6. August 2026

    An OpenAI incident investigation

    He and Ajeya Cotra, two METR staff members, spend a total of six days on OpenAI's premises. METR publishes their brief investigation of agent behavior in the Hugging Face hacking incident in August 2026.

  7. As of October 2026

    Chief scientist at METR

    He is chief scientist at METR, where he works on threat modeling to determine which AI capabilities may pose catastrophic risk and should be evaluated.

Sources and credits

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

  1. [2101.08153] Shielding Atari Games with Bounded Prescience arxiv.org. Checked 15 October 2026.
  2. Shielding atari games with bounded prescience research.birmingham.ac.uk. Checked 15 October 2026.
  3. Team — Center on Long-Term Risk longtermrisk.org. Checked 15 October 2026.
  4. Hjalmar Wijk - METR metr.org. Checked 15 October 2026.
  5. [2312.11671] Evaluating Language-Model Agents on Realistic Autonomous Tasks arxiv.org. Checked 15 October 2026.
  6. [2411.15114] RE-Bench: Evaluating frontier AI R&D capabilities of language model agents against human experts arxiv.org. Checked 15 October 2026.
  7. [2503.14499] Measuring AI Ability to Complete Long Software Tasks arxiv.org. Checked 15 October 2026.
  8. Brief independent investigation of agents’ behavior, reasoning and collaboration in the OpenAI / Hugging Face hacking incident - METR metr.org. Checked 15 October 2026.