Cloakpipe

Inspiration

We noticed that AI agents are getting really good at doing things, but there isn't much infrastructure to control what they're actually allowed to do. If an agent gets prompt injected or simply makes a bad decision, it can still call tools or APIs it shouldn't. We wanted to build a security layer that stays in control even when the model doesn't.

What it does

CloakPipe sits between AI agents and the outside world. Every tool call, API request, or agent action passes through it first, where policies decide whether to allow, block, or redact it. Every decision is also recorded in a tamper-evident audit log so developers can see exactly what happened.

How we built it

We built CloakPipe as an open-source proxy that intercepts outbound agent requests before they reach external services. The policy engine evaluates each request in real time and either allows, blocks, or modifies it based on configurable rules. Everything is logged to create a complete audit trail for debugging and security reviews.

Challenges we ran into

One of the biggest challenges was making policy enforcement flexible without adding noticeable latency. We also spent a lot of time designing a system that could work across different agent frameworks instead of being tied to a single ecosystem.

Accomplishments that we're proud of

• built a working open-source implementation
• integrated it into a production environment
• created deterministic policy enforcement outside the model
• added tamper-evident audit logging for every agent action
• designed it to work with different AI agent frameworks

What we learned

Building capable AI agents is only half the problem. As agents become more autonomous, having independent control and visibility over their actions becomes just as important. We also learned that security infrastructure has to be easy to adopt if people are actually going to use it.

What's next for Cloakpipe

Next, we want to turn the core enforcement engine into a standalone library, improve documentation, add integrations with popular agent frameworks, publish a formal threat model, and get external security reviews. Our long-term goal is to make external control and auditing a standard part of AI agent infrastructure.

Built With

• rust
• typescript