Inspiration

AI agents are being deployed in healthcare, finance, and enterprise settings, but traditional guardrails only check inputs and outputs. They miss what happens during the run. We wanted to make Computer and Browser Use Agents safe for regulated industries where a single unauthorized access is a compliance violation.

What it does

Sec0 SDK provides runtime governance for AI Agents in production. It monitors every action an agent takes, blocks unauthorized access in real-time, and generates tamper-proof audit trails. In this demo, a healthcare agent gets blocked when it tries to access mental health records it wasn't authorized to view.

How we built it

We built a TypeScript SDK with decorators that wrap agent and tool calls. Each action is checked against a pre-approved mandate. If the agent deviates, the action is blocked before it completes and optionally fix and replay the broken runs with baseline versions. Screenshots and logs are captured at every step for compliance evidence and to determine the exact path of risk.

Challenges we ran into

Instrumenting async agent loops without breaking execution flow. Designing mandates that are strict enough for security but flexible enough to be practical.

Accomplishments that we're proud of

Creating the most convenient and pluggable SDK including how fast we were able to build the entire infrastructure.

What we learned

Computer-Use Agents have the largest attack surface of any AI agent, they can see and interact with the entire system. Healthcare, Finance, and Insurance compliance requires proof of what didn't happen as much as what did.

What's next for Sec0 SDK

We are actively looking for 2-3 design partnerships (Cua if you're seeing this and for your enterprise Customers) to shape AI Agent Security and GRC for enterprises. Also, soon to be open-Source.

Share this project:

Updates