Independent testing and assurance for AI and the systems it runs.
Red team to blue team. We find how a system fails, then stay with your team until the fix holds.
What we do and why
AI is moving out of the chat window and into the systems people depend on.
Models now act on their own inside companies and critical systems. A failure there reaches people.
Most testing still looks at the model on its own.
We test it where it runs, against the harm that would matter most there.
Most red teams stop at the report. We stay with your team until the attack that worked fails on retest.
That is syntony: two systems tuned to the same frequency. Why the name
Live workshop, December 1. AI Red Teaming and Risk Governance, 1 to 4 p.m. ET, online. 15% off through November 15. Details and registration
What we do
We test AI models and agents, robotics and autonomy, quantum computing, space and defense systems, telecom and energy networks, and the public institutions that depend on them. Who we work with
- Red teamingAdversarial testing of models, agents and the systems they control.
- EvaluationsRepeatable tests built from what the red team found.
- BenchmarkingHeld-out sets and validated graders for the capabilities you care about.
- Governance advisoryControls and owners tied to the evidence, so decisions hold up later.
- Training and exercisesWorkshops and wargames for teams that build or oversee AI.
From red to blue
A red team finds how a system fails. A blue team fixes it and keeps it fixed. Most engagements end in between, with a report.
That gap is wide. In Cobalt’s penetration testing data, fewer than half of all findings get fixed, and for generative AI applications it is about one in five.
We stay through the fix. A finding closes when the attack that worked fails on retest, and that record is what evaluation, governance and assurance rest on.
Built by a team from
U.S. Department of War
Our red team
Our red teamers have worked for frontier labs, big tech, and government and international organizations.
When a system needs depth we do not keep in house, we bring in specialists from our Trusted Red Team Network.
About the networkResearch
Evaluation Cards, a paper on how AI evaluation results are reported, was selected as a spotlight at NeurIPS 2026. Nathan Heath and Dr. Wm. Matthew Kennedy are among its authors, with the EvalEval Coalition.
Read about the paperWhy Syntony
Syntony means being tuned to the same frequency. In 1889 Oliver Lodge showed it with two Leyden jars: charge one, and the other answers only when its circuit is tuned to match.
We took the name for the same reason. Testing helps only when it is tuned to the system in front of it.
Tell us what you need tested
We reply within two business days.