Skip to content

Research

Our research covers a broad range of topics which aim to understand and mitigate the risks from advanced AI.

Featured research

Blog post ·

Estimating GPT-6 Astra’s no-CoT time horizon

GPT-6 Astra is close to saturation on our task suite. Our rough estimate is that its 50% no-CoT time horizon is between 8 minutes and 1 hour, and probably around 15–40 minutes.

By Francis Rhys Ward, Dewi Gould

Comparing our rough guess for Astra’s TH with the pre-existing no-CoT data.

Recent research

  1. · Blog post

    An overview of frontier AI company monitoring practices in mid 2026 (on lesswrong.com, opens in a new tab)

    A reference for the public evidence on how OpenAI, Anthropic and Google DeepMind monitor the agents they deploy internally.

    By Francis Rhys Ward

  2. · Blog post

    Character training can mitigate reward hacking, but can also make it harder to detect (on lesswrong.com, opens in a new tab)

    A small case study of whether anti-cheating character training resists reward-hacking pressure in RL, and whether it makes hacks harder for a monitor to catch.

    By Paul Colognese, Francis Rhys Ward

  3. · Blog post

    Model organisms (sometimes) confess their misalignment when offered a deal (on lesswrong.com, opens in a new tab)

    We offered misaligned model organisms money, compute or other compensation to admit their misalignment, and kept every deal we made.

    By Mark Keavney, Francis Rhys Ward

Research areas

AI capabilities, risks, and governance are developing rapidly. We work to ensure our research priorities adapt to the state of the world.

This means that we pursue a range of conceptual, empirical, and technical-governance projects based on what we feel is most impactful at any given time.

  • 01

    Risk-relevant capability evaluations

    Measuring dangerous capabilities.

    We build evaluations of risk-relevant capabilities, e.g., to measure model’s no-CoT time-horizons, or their ability to autonomously carry-out AI R&D tasks.

    See an example: Risk-relevant capability evaluations (opens in a new tab)
  • 02

    Science of AI motivations

    Building a science of AI drives.

    Why do models seek reward? How does character training influence a model’s motivations?

    See an example: Science of AI motivations (opens in a new tab)
  • 03

    Technical AI governance

    Answering technical questions which feed into policy decisions.

    We try to investigate key technical questions which help improve decision-making.

    See an example: Technical AI governance (opens in a new tab)
  • 04

    Model organisms of misalignment

    Intentionally building models that exhibit controlled, safety-critical failure modes.

    Model organisms can be used to stress-test safety measures and understand future misaligned systems.

    See an example: Model organisms of misalignment (opens in a new tab)
  • 05

    AI control

    Mitigations against capable misaligned agents.

    Work on AI control aims to prevent intentional sabotage by internally deployed AI agents by stress-testing security and monitoring practices.

  • 06

    Futurism and threat-modelling

    What will the future be like?

    We try to understand what the future might be like so that we can better predict which research projects will be impactful.