Intelligence should scale. Human control should never disappear.
AI systems can now act faster than the safeguards built to oversee them. We work on the foundations that keep people in charge.
What we believe
As AI systems gain the ability to act independently, humanity must retain the ability to understand, constrain, verify, and override their actions.
Safety should be a property of the architecture, not a promise made by the developer. A promise can fail under pressure; an architectural boundary holds.
Vision and mission
Vision
Increasingly capable AI that remains safe, verifiable, accountable, and subject to human control.
Mission
Advance the technical foundations of AI safety through open research, reproducible evaluations and accessible tools.
Five principles, each with an open question.
A principle is only useful if it can be tested.
- 01
Human Control
People keep the authority to set boundaries, intervene, and revoke permissions.
Open questionHow precisely can authority be delegated, and how quickly can it be withdrawn?
- 02
Verifiable Safety
Safety claims rest on reproducible experiments and published methods.
Open questionHow do we measure what a safeguard stops and what it breaks, together?
- 03
Security by Design
Protection lives in the architecture, not in instructions to the model.
Open questionWhich controls must sit outside the model to hold under adversarial input?
- 04
Open Knowledge
Foundational safety tools and methods are open to everyone.
Open questionWhich protections are basic enough that no one should have to pay for them?
- 05
Accountability
Autonomous actions stay traceable and open to independent review.
Open questionWhat must a record contain for someone else to verify a decision?
What we hold our own work to.
We are run by a security company. These rules keep that visible.
- 01
Status on every claim
Published, ongoing or proposed — never blurred.
- 02
Method before conclusion
Results ship with how they were measured, failures included.
- 03
Interests disclosed
When we evaluate our own products, we say so first.
- 04
Open by default
Tools and methods are released under open licenses.