Estes

🚀 Estes — Securing the AI Agent Supply Chain

Live Demo: https://www.verylowcortisol.vip/

A pre-install auditor and runtime redaction wrapper that prevents credential leaks and tool-output poisoning in AI agent ecosystems (OpenClaw, Antigravity, Claude Code, watsonx Orchestrate-style systems).

🧠 The Problem

Recent research (April 2026) auditing 17,022 public AI agent skills found:

3.1% leak real API keys, OAuth tokens, or passwords 73.5% of leaks come from leftover print() / console.log() debug statements Agent frameworks automatically inject stdout into LLM context windows

👉 Once secrets enter the context, they can be extracted with a simple follow-up prompt—no hacking required.

This creates a silent supply-chain vulnerability affecting any team using third-party agent skills. Major platforms (e.g., enterprise orchestration systems) have already flagged this as a critical risk.

🛠 What Estes Does

🔍 Pre-Install Auditor

Upload a .zip or GitHub repo → get a risk score in seconds.

Static scan: Regex + entropy detection for: API keys, OAuth tokens Wallet keys, SSH keys Cloud/database credentials AST analysis (Python): Tracks sensitive data (os.environ, wallet factories) Detects flows into print(), logs, or network calls

Output:

Severity-ranked findings Line-level explanations Fix suggestions JSON export ✅ “Download Fixed” → auto-patched safe version + change log

⚡ Runtime Redaction Wrapper

FastAPI middleware that sanitizes tool output before it reaches the LLM.

Sub-40ms latency Drop-in for agent frameworks Prevents secrets from ever entering context

📜 Compliance & Audit Hooks

Every redaction creates a structured log:

Rule triggered Severity Skill affected Action taken

→ Ready for enterprise governance and audit pipelines

🏗 Architecture User → Agent Framework ↓ Third-party skill ↓ Tool output (stdout/result) ↓ Estes Wrapper (FastAPI) ↓ Detection → Redaction + Logging ↓ Clean output → LLM context 🧪 Demo Flow Open https://estesss.vercel.app Upload a repo or demo skill Watch real-time risk scoring (<2s) Inspect flagged leaks and fixes Download patched version Test runtime redaction in-browser ⚙️ Tech Stack Backend: FastAPI Frontend: Vanilla JS + Tailwind Analysis Engine: Python AST + regex + entropy heuristics Runtime: Uvicorn middleware Deployment: Vercel (serverless Python) 🧩 Key Innovation No LLM dependency → fast, deterministic, privacy-safe Combines static + runtime protection (rare in current tools) Targets a new vulnerability class unique to agent systems Works across multiple agent frameworks 🗺 Roadmap TypeScript / JS AST support CLI (estes scan, estes start) Policy-as-code + allowlists Chat “Shield” before sharing/exporting conversations Enterprise governance integrations 🎯 Who It’s For Developers building with AI agents Teams using third-party skills/plugins Enterprises deploying agentic AI at scale 🏁 Built For

Hook ’Em Hacks 2026 — Security in an AI-First World (IBM Track)