Our research team builds and tests models that detect AI across audio, images, video, and agent behavior, so trust endures as AI reshapes how we communicate, work, and decide what’s true.
01
Adaptive
Treating every advance in AI as a new problem, so detection evolves alongside the technology it’s built to catch.
02
Expansive
Going wherever AI goes, from voice to images, video, and behavior, and on to whatever comes next.
03
Responsible
Building with respect for security, identity, and privacy, because the data being analyzed belongs to real people.
04
Global
Monitoring threats and regulations worldwide, because AI misuse crosses borders and no region faces it alone.
Lab conditions aren’t real conditions. Calls get compressed, platforms re-encode uploads, and rooms add echo. So our research team built Proteus, a framework that automatically chains everyday audio processing to find the combinations that can flip our own detector’s verdict. Every weakness it uncovers goes back into training, and Proteus runs again to confirm the fix. It’s a closed loop between finding a blind spot and closing it.
17,405 processing chains tested
~110 audio variants
Presented at ISSMAD 2026
Our detection models are built to help answer the fundamental questions organizations need to restore agency in a changing world.
The first world model for detection: combines pattern matching and training with a world model architecture that recognizes deviations from physical reality so it can catch fakes from models it has never seen.
Extends detection from what AI creates to how AI acts, identifying AI agents by their behavior, allowing organizations to control access and determine a course of action.
Turns detection results into plain-language explanations, so people can understand and act on what the model found.
Scores content against the fraud categories each organization defines, so trust decisions reflect the threats that matter to them.
Embeds imperceptible watermarks across audio, images, and video, so provenance travels with the content.
Enrolls a person’s voice and likeness, then verifies them against it, so you can confirm who’s real, not just flag what’s fake.
