Preparing for the
intelligence
explosion.
A technical AI safety lab working on data poisoning research and red-teaming, in preparation for the intelligence explosion.
Mitigating the risks of Recursive Self Improvement (RSI).
We are a technical AI safety lab focused on the security and safety gaps that become increasingly consequential as AI development becomes more automated. Our agenda is built in collaboration with researchers across the AI safety ecosystem — frontier labs, nonprofits, and government organisations — to identify and address the most pressing risks. Our initial focus areas are below, and will grow with the field.
- 01
Stealthy and Persistent Data Poisoning Attacks
Understanding how training data can be manipulated to produce harmful, hidden, or persistent behaviours, and developing evaluation and mitigation strategies that scale with automation.
- 02
Model Organisms
Building small, tractable model organisms of misalignment so failure modes can be studied and measured before they appear at frontier scale. Self-propagating secret loyalties — hidden objectives that survive successive generations of training, evaluation, and deployment — are one example.
- 03
Data Systems Red-Teaming
Stress-testing the data pipelines, curation tools, and evaluation infrastructure that frontier labs rely on to find failure modes before they are exploited.
Build the future with us.
Knowledge is shared openly, ambition is matched by rigor, and every researcher is empowered to work on what matters most. If you thrive on hard problems and care deeply about safety, we'd love to hear from you.
Apply →