Lilac

We’re three young programmers and engineers with different backgrounds, but a shared interest in agentic systems and feedback loops.

We built Lilac to explore a simple idea: when an AI agent fails a security test, that failure should become a permanent regression test, and the defense should improve automatically.

Our demo uses a customer-operations agent with tools for support, refunds, CRM data, and email. A Red Agent looks for unsafe attack paths, a Blue Agent proposes defenses, and a deterministic verifier checks whether the attack really worked and whether normal workflows still function.

One of the biggest challenges was working with unfamiliar tools, especially Pomerium and MCP-style tool policies. Getting the policy and enforcement model right took time, but it also pushed us toward a stronger design: the model can be wrong, while the system still blocks dangerous actions.

We also learned that the hard part is not generating attacks. It is proving that an attack succeeded, avoiding over-broad defenses, and keeping every result organized so the system improves over time.