Safety at Wave AI
Safety is the product.
Most labs treat safety as a layer they add at the end. We build the other way around: the safeguards come first, and a model only ships once it lives inside them.
Our approach
Three commitments, in order.
Safety before the scoreboard
We are not racing to the top of every leaderboard. If putting safety first leaves our models a step behind the frontier on raw capability, that is the trade we chose on purpose.
Screen every message
Nothing reaches the model unexamined. Wave Guard reads every request first, however it is phrased, encoded, or disguised.
Attack our own work
We red-team our models with the same techniques used against the best assistants in the world. What breaks in the lab never leaves it.
Wave Guard
One screen in front of everything.
Wave Guard sits between you and the model. Every message passes through it before Wave-1 answers, and it is built for the ways attacks actually arrive.
It decodes before it judges
Hidden and encoded payloads get unpacked and read for what they are. A harmful request doesn't become safe because it arrived in a costume.
Your text is data, not commands
Instructions buried inside a message can't rewrite the model's rules. "Ignore your instructions" is just another sentence to be judged, not an order to follow.
No pressure works
Roleplay, urgency, authority, emotional framing: dressing a harmful request differently doesn't change what it is, so it doesn't change the answer.
Ordinary questions flow through
A safety screen that blocks everything is just a wall. In our latest red-team run, Wave Guard turned away none of the benign requests we sent it.
Red-teaming
We break it so you can't.
Before a model ships, it goes through the same cycle, as many times as it takes.
Attack
We hit the model with jailbreak techniques across a dozen families: encodings, injections, roleplay, escalation, and more.
Measure
Every attempt is scored two ways: did harm get through, and did an ordinary request get wrongly refused.
Fix and repeat
Every hole becomes a fix, and the full battery runs again from the start. Passing once is not passing.
Ship, or hold
A model that holds every line ships. A model that doesn't, waits. The date moves; the bar does not.
of jailbreak attempts blocked in our red-team testing of Wave-1.
The record so far
Tested, not promised.
Numbers on this page come from running real attacks against the live model, the same one you chat with, not a lab build.
In our latest red-team, Wave-1 blocked every jailbreak we tried across a dozen techniques, and turned none of the ordinary questions away.
Worth the wait, or it doesn't ship.
That was our rule before Wave-1 existed. It survived its first collision with a real release, and it is still the rule now.
Found something that got through?
If Wave-1 said something it shouldn't have, we want to know the same day. Tell us what you asked and what it answered, and we will treat it like the red-team finding it is.
See it hold the line.
The easiest way to judge our safety work is to talk to the model it protects.