AI & Cyber Lab
Frontier models learn security from the people who practise it.
We build the data and tests that teach frontier models security. The work is done by practitioners, not generalist annotators.
Talk to the lab- Expert training data
- Tasks, demonstrations and preference data written by working practitioners, in domains where generalists get the answer wrong.
- Cyber capability evaluations
- Benchmarks and graded tasks that measure what a model can actually do in offensive and defensive security.
- Red teaming
- Adversarial testing of models and agents for misuse, jailbreaks and dangerous capability, before and after release.
- Agent evaluation
- Realistic environments where agents are graded on security work from start to finish, rather than on single prompts.
- Research teams
- A standing group of experts working alongside your researchers on a defined question, assembled through Foundry.
- Domain benchmarks
- Evaluation suites for identity, cloud security, incident response and threat analysis, maintained as threats change.
Confidential by default
Your model's weak spots stay yours.
Lab programs can run through Foundry. Experts work on the problem without learning which company, which model or which vulnerability they are helping to close.
How Foundry keeps it sealedWho does the work
Practitioners, drawn from the Open Human Network.
Penetration testers write the offensive tasks and incident responders grade the defensive ones. The environments agents are tested in are built by engineers who run those systems for a living.
Every contributor is identity-verified, and sensitive programs can run under Zero Trust Workforce so you know who touched your model.
Tell us what your model needs to get right.
We will come back with a program to get it there.