Renaissance Labs

Atlas / Europe

Institute · London, United Kingdom

Apollo Research

London-based AI safety organisation that evaluates frontier models for scheming and publishes open evaluation and interpretability work.

AI safetyevaluationsinterpretability

London, United Kingdom

Founded
2023
Type
Institute
Based in
London, United Kingdom
Key papers
2
Featured projects
7

About

Apollo Research was founded in 2023 and is led by CEO and co-founder Marius Hobbhahn. Its stated goal is to secure frontier AI systems across development, deployment and governance, with a focus on scheming, where advanced systems covertly pursue misaligned objectives.

In 2023 it presented at the UK AI Safety Summit and advised the Frontier AI Taskforce. In 2024 it tested OpenAI's o1 before release and published early evidence of in-context scheming in frontier models; in 2025 it ran evaluations for major labs and worked with OpenAI on anti-scheming interventions.

In 2026 the organisation became a Public Benefit Corporation, opened a San Francisco office and began building Watcher, an agent monitoring tool. It publishes code on GitHub, including end-to-end sparse dictionary learning.

Mission and ambitions

Apollo aims to reduce the risk of deceptive and misaligned AI through evaluations, interpretability research and governance advice. It works with developers and governments.

Known for

  1. 01

    Scheming evaluations of frontier models

  2. 02

    End-to-end sparse autoencoders

  3. 03

    Pre-deployment testing with OpenAI

  4. 04

    Loss of control playbook

Milestones

  1. 2023

    Presented at AI Safety Summit

    Apollo identified scheming as a serious risk and presented at the UK AI Safety Summit.

  2. 2024

    End-to-end SAEs

    Released training code and weights for end-to-end sparse dictionary learning (arXiv 2405.12241).

  3. 2024

    o1 pre-deployment testing

    Partnered with OpenAI to test o1 before public release.

  4. 2024

    In-context scheming evaluations

    Published evaluations showing frontier models can scheme in context (arXiv 2412.04984).

  5. 2025

    Anti-scheming work

    Partnered with OpenAI on anti-scheming interventions and wrote a loss of control playbook for governments.

  6. 2026

    Public Benefit Corporation

    Became a PBC, opened a San Francisco office and began building Watcher.

Open source

GitHubgithub.com/apolloresearch
Hugging Facehuggingface.co/apollo-research

Research papers

Notable releases

Openness

Research code, SAE weights and datasets are public; its evaluations of frontier models are published as papers.

People

Listed from the lab's own pages and public profiles.

Contact and community

Contact through the website.

Sources

Compiled from the lab's public pages, Hugging Face and GitHub. Last checked 2026-11-17. Something wrong or out of date? Tell us on Discord.

Labs nearby