GitHub Repository

A causal firewall for AI agents: blocks multi-step tool-call chains that leak data, even when every call is individually allowed.

1 starsRust

A Firewall for AI agents with auditing

Name: A Firewall for AI agents with auditing
Availability: InStock
Author: beebeeVB

by beebeeVB·Jun 29, 2026·2 points·0 comments

Visit Project View on HN

AI Analysis

●●●BangerWizardryBig BrainZero to One

Causal DAG tracking catches multi-step exfiltration that per-action security checks completely miss.

Strengths

•Trajectory-based enforcement blocks legal-in-isolation action chains that leak data across steps.
•Taint mechanism tracks data flow from sensitive sources through summarization to forbidden sinks.
•Sealed-commitment graphs with HMAC prevent runtime tampering; Rust delivers sub-5ms latency.

Weaknesses

•Early project with minimal stars—real-world agent attack patterns still unproven at scale.
•MCP-only integration limits adoption to agents using Model Context Protocol.

Post Description

Hi all,

As there are more and more agents in the internet; Security is going to be a big problem. Currently, the problem is solved using a LLM to guard Agent but this creates the problem of hallucination and latency, so I coded a firewall in rust that runs under five miliseconds. This works by creating a plan and enforcing the plan; for per action call, this enforces using the Model context protocols list and for sequence it tracks every single tool call and data flow; there is also a taint mechanism where if the agent reads something outside of the user context, it flags and adds more security mechanism. It works by using a DAG.