heisty-agents
A security agent that make sure everything running on behalf of the coding agent owner is safe and secure.
Core trick
Tree-sitter parses each Bash command before execution, an LLM judge reviews survivors, and escalations post to Slack.
Description
A security agent that make sure everything running on behalf of the coding agent owner is safe and secure.
The security agent uses a defense in depth approach powered by Gemini Argon 4
Tier 1 Tree sitter Abstract Syntax Tree:
its the foundation of using parsing to check against bash or CLI commands thats ran by the model
Tier 2: Argon 4 as a judge:
Anything that passes tier 2 gets filtered by Argon 4 using LLM as a judge approach. If it does not pass through tier 2
Tier 3: Escalates to the human (Human in the loop
it sends a webhook to slack channel with the LLM analysis and attack path that alerts the human.
cyberleo17-VGS/argon-hack-project-heist★ 00 forksPythonCreated Oct 25, 2026 · last push Oct 25, 2026
README excerpt
# 🛡️ Gemini CLI SecOps Guard (3-Tier MDR Pipeline)
**A Defense-in-Depth Security Operations (SecOps) guardrail for [Gemini CLI](https://github.com/google-gemini/gemini-cli) that intercepts, analyzes, and blocks dangerous Bash commands before they execute.**
---
## 💡 What is this project?
When AI coding agents are given access to a terminal (`bash`), they can accidentally—or through prompt injection—delete files, read sensitive credentials, or exfiltrate data.
**SecOps Guard** hooks directly into Gemini CLI’s `BeforeTool` lifecycle event. Every time the AI agent attempts to run a shell command, SecOps Guard pauses execution and runs the command through a **3-Tier Managed Detection & Response (MDR) pipeline**:
[Gemini CLI Agent proposes Bash Command]
│
▼
┌─────────────────────────┐
│ 🪝 BeforeTool Hook │
└─────────────────────────┘
│
▼
┌───────────────────────────────────┐
│ TIER 1: Tree-Sitter AST + YAML │ ──(Blocked Binary / Syntax)──┐
│ (<5ms deterministic static check) │ │
└───────────────────────────────────┘ │
│ (Passed) │
▼ ▼
┌───────────────────────────────────┐ ┌───────────────────┐
│ TIER 2: Gemini-4-Argon AI Judge │ ──(Malicious)────► │ TIER 3: Slack MDR │
│ (Semantic in