Lilith (BadCompany, 2025 to 2026)
At BadCompany I was responsible for technology, and we built Lilith, a security product for AI agents. If an agent can call tools, it will eventually do things it has access to. Thus we need a way to define these boundaries in a finegrained and seamless (non annoying) way. So we built a robust and scalable solution sitting below the agents.
Lilith has two parts:
1. a policy engine providing an expressive, finegrained and interoperable interface for defining policies for various tools and agent integrations (some properties of the policy can be formally verified by Lilith to be obeyed by it before runtime!);
2. a kernel level policy enforcer which is unbypassable from userspace (where the agents are running), observes all actions, manages secrets and is invisible to the AI agent;
We took inspiration from many products and solutions using kernel extensions (mainly eBPF), like Cilium, and also many userspace solutions.
We are now in a testing phase with the largest bank in Central and Eastern Europe for a full rollout to on the order of 4,000 developers. I led the technology side from the policy language through kernel enforcement and the path to production.