Agent Leak Guard

Inline secret leak scanner, redactor, CLI, and git hook for AI coding agents.

Scans content an AI coding agent is about to write, commit, or output — files, diffs, terminal output, logs — for accidentally leaked API keys, tokens, or credentials, and blocks or redacts them before they land anywhere.

Installation

npm install agent-leak-guard

Quick Start

# Scan a file
npx agent-leak-guard scan path/to/file.js

# Scan from stdin
cat file.js | npx agent-leak-guard scan --stdin

# Install pre-commit hook
npx agent-leak-guard install-hook

Library Usage

import { scan } from 'agent-leak-guard';

const result = scan('AWS_KEY=AKIAIOSFODNN7EXAMPLE', {
  mode: 'block', // warn | block | redact
});

console.log(result.hasFindings); // true
console.log(result.isBlocked);   // true
console.log(result.findings[0].rawMatch); // AKIA****MPLE (never the full secret)

Response Modes

| Mode | Behavior | |------|----------| | warn | Flag the finding, redact in output, allow action to proceed | | block | Prevent write/commit/output, show redacted finding and reason | | redact | Replace detected secret with [REDACTED:<type>] and proceed |

Configuration

Create .leakguard.json in your project root:

{
  "defaultMode": "block",
  "entropyThreshold": 4.5,
  "entropyMinLength": 20,
  "allowlist": {
    "strings": ["AKIAIOSFODNN7EXAMPLE"],
    "regexes": ["^TEST_", "^EXAMPLE_"],
    "files": ["**/node_modules/**", "**/*.min.js"]
  },
  "rules": {
    "aws_access_key": { "enabled": true, "mode": "block" },
    "jwt_token": { "enabled": true, "mode": "warn" },
    "entropy": { "enabled": true, "mode": "warn" }
  }
}

Integration with agent-permit

Use agent-leak-guard as an inline check in your agent middleware:

import { scan } from 'agent-leak-guard';

async function onFileWrite(content: string) {
  const result = scan(content, { filePath: '/path/to/file' });
  if (result.isBlocked) {
    throw new Error(`Secret leak blocked: ${result.summary.blocked} finding(s)`);
  }
  return result.redactedContent;
}

Detectors

• AWS Access Keys (AKIA/ASIA/ABIA/ACCA)
• OpenAI API Keys (sk-...)
• Anthropic API Keys (sk-ant-api...)
• GitHub Personal Access Tokens (ghp_, gho_, ghu_, ghs_, ghr_, github_pat_)
• Stripe API Keys (sk_test_, sk_live_, rk_test_, rk_live_)
• Private Key Blocks (-----BEGIN ... PRIVATE KEY-----)
• JSON Web Tokens (JWT)
• Database Connection Strings with credentials
• Slack Tokens (xoxb-, xoxa-, xoxp-, xoxr-)
• Google Cloud API Keys (AIzaSy...)
• High-Entropy Strings (fallback for unknown tokens)

Patterns reference open-source standards such as gitleaks and trufflehog.

Never Leak the Leak

All findings are redacted before exposure. The rawMatch field contains only a masked preview (e.g. AKIA****MPLE). Full secret values are never logged, returned, or persisted.

License

MIT