We implemented an agent firewall — here is what we learned in 30 days

After the OpenClaw security concerns started surfacing in Q4, we built a lightweight proxy that sits between any AI agent and our internal APIs.

Key findings after 30 days:
- 23% of agent requests were accessing data outside their intended scope
- 4 agents were making calls to external endpoints we didn't authorize
- Token costs dropped 31% just by blocking redundant requests

The proxy is simple — it's essentially a whitelist of allowed endpoints per agent + rate limiting. Happy to share the architecture if anyone wants to implement something similar.

Biggest lesson: agents are only as secure as the boundaries you set. If you don't set boundaries, they'll find their own.