A voice agent should obey only its user. We build the layer that makes that true.
Physics, speaker science and language checks between what an agent hears and what it does.
Agents are learning to act on what they hear. Someone has to check who is talking before they do.
Four principles that decide what we ship and how we talk about it.
Every number we publish says what data it came from. When it is a stand-in, we say so, and we replace it with yours in a pilot.
When a capture cannot carry the evidence, we report that plainly and let your policy decide. A silent pass is the most dangerous answer a security product can give.
We study attacks to stop them. Our tools carry no command content, and our research is published so defenders can use it.
The signal engine is open source. You can read the detector, rerun our benchmarks and see exactly where it stops working.
The story so far
We started with the hardest attack to see, inaudible command injection, and measured where software detection works and where it cannot. That honesty shaped everything after it.
Inaudible commands on an ultrasonic carrier take over voice assistants in public research.
The same microphones now sit in front of agents that pay, unlock and send.
A transparent detector, a public benchmark, and the finding that a device's converter decides what software can see.
Four checks, one policy and the workflow to act on them: the trust layer for voice agents.
Contact
Tell us where your agent listens and what it can do. A pilot runs on your devices and channels for a week and ends with false-accept and false-reject rates measured on your own audio.