Who Can You Trust?

2026-08-11T11:00:00-04:00

These days, it’s getting harder to know not just which AI you can trust, but whether you can trust AI at all. In the past two weeks, agents from half the leading AI companies suffered major failures (agents breaking out of sandboxes, deepfake photos as maps entries). And we’re still in the opening act of what many call the “AI takeoff.”

This is all rather depressing. Then again, the point isn’t really that you can trust AI, but that you need humans you can trust to steer your ship as close to true north as possible.

This “true north” is what many people call “AI assurance.” The Partnership on AI describes AI assurance as “the process of measuring, evaluating and communicating the trustworthiness of AI systems.”

In many ways, that’s our goal at Syntony. We aim to provide our clients with continuous AI evaluation that influences good governance and furthers AI assurance. Most of the time, the first step in our process is red teaming (“interactively testing AI models to protect against harmful behavior”).

Red teaming is the keystone of AI assurance, so we put a lot of effort into making sure we get it right. That’s why our red-teaming process is a bit different, and what makes it needed at a time when trusting the models we use is at a crossroads. Four principles guide our philosophy:

Evaluate proactively, not reactively. We help organizations anticipate risks from loss of control and CBRNE misuse to sycophancy and disinformation. As much as possible, we prefer to fireproof organizations against AI, not just put out their fires. This also means we design multi-phase campaigns, not two-week sprints.
Assess a system of risks, not an isolated one. Unsafe AI doesn’t manifest in a siloed pattern. Threat actors often run AI-generated information campaigns while leveraging AI-embedded malware or scams, and terrorist groups chase CBRN recipes while also using deepfake photos for propaganda. This requires a holistic, systems-thinking approach to assessing AI red-teaming risks, à la Donella Meadows. Syntony’s risk-mapping approach, supported by its forthcoming software, Resonance, fills this gap.
Red-teaming by humans, for humans. We test across major languages, modalities, and risk taxonomies. While automated testing has its place, we believe that adversarial evaluation of AI models relies first and foremost on expert human insights. Our Trusted Red Team, an elite network of external subject-matter experts who’ve spent years breaking some of the top AI models in the world, ensures that customers are getting the very best “human touch” available.
Deliver lasting guidance to enable continuous evaluation, not constant correction. We end every red-teaming engagement in person with the client, running onsite, multi-day workshops to present red-teaming findings and map their implications for governance and assurance. This allows our team, including our experts, to personally demonstrate AI risks to the client and convey concrete safety recommendations to shape good governance, whether regulatory alignment or policy frameworks, that last beyond a single campaign.

We don’t claim that when we’re done, an AI model will never act unsafely again. There’s no “Syntony-approved” rubber stamp. Continuous adversarial evaluation can and will be necessary to address persistent or emerging model risks.

What we will claim is that we offer a uniquely interdisciplinary, human-centered approach allowing a client to wholly evaluate their AI risks and consider their opportunities. Moreover, we believe that by viewing our evaluation process as part of an interconnected chain with governance and assurance¹, we can come closer to closing the trust gap between humans and machines.

It’s clear that AI is here to stay, and even though we have only so much control over how it’s developed and released, we can do our part to make it as safe as possible, for the sake of both those within our orbit and humanity as a whole.

Syntony also offers a range of products and services focused on governance and assurance beyond its red teaming campaigns ↩

Syntony Insights

Who Can You Trust?