Wrynx Joins the Open Secure AI Alliance
Why we joined
The Alliance starts from a simple premise: defenders need security tools they can inspect, adapt and run on infrastructure they control. As AI moves into production software, agents and critical infrastructure, closed and opaque defenses aren't enough. Security teams need to understand how their protections work, verify them and change them as models and threats evolve.
That premise is the reason Wrynx exists.
Most AI security today works at the edges of a model. It filters what goes in and checks what comes out. Those layers matter, but they only see the surface. Much of what a model is actually doing happens inside it, in the internal representations it builds as it processes each input.
Security from inside the model
Wrynx builds latent space probes: lightweight models that read an AI system's internal activations and flag the behaviors you care about. Because probes look at what the model is computing rather than only at its final output, they can surface signals that surface-level filters miss.
Probes are also a naturally open defense. They're small, auditable and easy to reason about. You can see exactly what a probe was trained on, test it against your own data and run it wherever your model runs. That fits the Alliance's vision of defenses that are inspectable, portable and under the defender's control.
Our work spans three areas:
- Pre-built probes for several popular open-weight models, ready to deploy.
- Open-weight models with probes already fitted, so teams can start with monitoring built in.
- A platform for building your own probes: harvest activations from your models, train simple probe models on the behaviors that matter to your organization, and run them on your own stack.
What's next
AI security is a shared challenge, and no single company will solve it alone. The Alliance gives organizations a neutral home, under the Linux Foundation's governance, to share what they learn, test systems together and turn findings into common defenses.
We're looking forward to contributing what we've learned about model internals and probe-based monitoring, and to learning from the broader community of defenders, researchers and builders in the Alliance.
If you're working on securing open-weight models, or want to see what probes can reveal about the models you run, we'd love to hear from you. Get in touch.
Learn more about the Open Secure AI Alliance at secureaialliance.org.