Security for AI,and the software behind it
What We Do
Three lines of work, one team. Most engagements draw on more than one.
Securing AI Systems
Model and artifact signing, provenance across training and inference pipelines, key management for AI infrastructure, and permission and data-egress boundaries for agents that act on real systems.
Software Development
We design and build production software — applications, APIs, services, and the infrastructure under them — with a small senior team. Fixed scope, working code, and the tests and documentation to hand it over cleanly.
Cryptography & Security Engineering
Data sealed at rest and opened only to a person who is present, threat modeling, secure design and code review, PKI and key custody, and post-quantum migration when you need it. This is the bench the other two lines are built on.
AI Changed What You Have to Defend
A model reads like a person, at machine speed, and never loses interest. Models, pipelines, and agents are now part of your attack surface — and most of them were deployed faster than the controls around them.
Models Are Supply Chain
Weights and adapters pulled from a hub are executable artifacts with no signature, no provenance, and no record of what changed. You would never ship a dependency this way.
Agents Hold Real Permissions
An agent with tool access has credentials, network reach, and the ability to act. Prompt injection turns that reach into an exploit path unless the boundary is enforced outside the model.
Code Ships Faster Than Review
AI-assisted development multiplies output, and review capacity has not kept pace. The gap between what is written and what is understood is where the next class of vulnerabilities lives.
A Model Reads Whatever It Is Given
An AI with access to a directory reads it the way a person would — all of it, instantly, without getting bored. Access control written for humans assumed a reader who would not open every file. That assumption is gone.
Where We Work
Teams shipping AI into environments where a mistake is expensive — regulated, safety-critical, or simply too visible to get wrong.