Safety
Power without
accountability
is a liability.
We build AI that can reason about the world's hardest problems. That kind of capability demands a matching level of rigor about what can go wrong — and how we prevent it.
How we think about it ↓Last reviewed — June 2026
v2.4.1
Threat model
Four layers. All active.
What the model does
We define explicit behavioral constraints during training — not just as filters on output, but as internalized tendencies. The model should refuse harmful requests because it understands why, not because a regex caught the word.
99.7%
refusal accuracy on red-team evals
Our principles
What we commit to.
Hover each principle to read our reasoning.
Understand before deploying
Failures are features
No single point of trust
Transparency over optics
Users are people
System checks
Current status.
Our commitment
We publish our safety methodology, our eval results, and our failures. If you find a vulnerability in Caether's systems or models, we want to hear from you — before anyone else does.