AIHound is an AI Assistant credential and secrets scanner that detects exposed API keys, OAuth tokens, MCP server secrets, and session credentials across 29 AI tools on Windows, macOS, Linux, and WSL. Beyond one-shot scanning with terminal, it offers a watch mode for continuous monitoring that alerts on new, changed, or escalated credentials in real time.

AIHound includes an MCP server mode that lets AI assistants like Claude Code scan for and remediate credential issues programmatically.

AIHound can export to SpectorOps' BloodHound. Scan results export as OpenGraph JSON that can be ingested into BloodHound to visualize attack paths, showing compromised credential chains through MCP servers, AI services, and datastores. I've included 29 pre-built Cypher queries for blast radius analysis, same-secret detection, and lateral movement mapping.

This is a security research tool. Credentials are redacted by default so output is safe to share in reports and screenshots.

Get scanning in under 2 minutes.

PyInstaller Precompiled .exe version can be found Here

Go Precompiled .exe version can be found Here

AIHound can be run four ways: from Python source, using the Go runtime. as a compiled Go binary, or as a standalone Windows executable.

Full Documentation located Here

Prerequisites

• Python 3.10 or higher
• That's it. No pip install required for basic scanning.

Step 1: Get AIHound

git clone https://github.com/netwrix/AIHound.git
cd AIHound

Step 2: Run Your First Scan

python3 -m aihound

You'll see output like this:

╔══════════════════════════════════════════════════════════════╗
║          AIHound - AI Credential & Secrets Scanner           ║
╚══════════════════════════════════════════════════════════════╝

Platform: wsl
WSL detected - scanning both Linux and Windows credential paths

Tool             Credential Type        Storage      Location                            Risk
-------------------------------------------------------------------------------------------------
Claude Code CLI  oauth_access_token     plaintext... ~/.claude/.credentials.json          CRITICAL
                   Value: sk-ant-oat01-Z...eAAA
Claude Code CLI  oauth_refresh_token    plaintext... ~/.claude/.credentials.json          HIGH
                   Value: sk-ant-ort01-r...ygAA

Summary: 2 findings | 1 CRITICAL | 1 HIGH

All secret values are automatically redacted. The tool is read-only and doesn't touch your credentials.

Step 3: Get More Detail

Add -v for verbose output — shows file permissions (with human-readable descriptions), ownership, expiry times, and notes:

python3 -m aihound -v

Claude Code CLI  oauth_access_token     plaintext... ~/.claude/.credentials.json          CRITICAL
                   Value: sk-ant-oat01-Z...eAAA
                   Note: Expires: 2026-03-09 23:30 UTC
                   Perms: 0777 (world-writable, world-readable, DANGEROUS) Owner: ull

Step 4: Generate Reports

HTML Report

Creates a self-contained HTML file with the AIHound banner, dark theme, and color-coded risk table:

python3 -m aihound --html-file report.html

Open report.html in your browser. Great for sharing with your team or including in assessments.

JSON Report

For automation, pipelines, or feeding into other tools:

# Write to file
python3 -m aihound --json-file report.json

# Pipe to stdout
python3 -m aihound --json | jq '.summary'

BloodHound Attack Path Graph

Export to BloodHound CE for interactive attack path visualization — see how credentials chain together across tools, services, and data stores:

python3 -m aihound --bloodhound aihound-bloodhound.json

Then upload aihound-bloodhound.json to BloodHound CE (v9.x) via Quick Upload or Data Collection > File Ingest.

First time? Register custom node types and saved Cypher queries (once per BloodHound instance):

register_ai_nodes.py script located in docs folder.

python3 docs/register_ai_nodes.py -s http://<bloodhound IP>:8080 -u admin -p <password>

This registers 14 custom node kinds with icons and imports 29 saved Cypher queries into BloodHound's Saved Queries panel. Use --reset to re-register, --unregister to remove everything, or --no-queries to skip query import.

Example Cypher queries (also available in Saved Queries after registration):

// Show the full credential graph
MATCH path = (a:AIHound)-[r]->(b:AIHound) RETURN path

// Blast radius from critical credentials
MATCH path = (c:AICredential)-[*1..4]->(target)
WHERE c.risk_level = "critical"
RETURN path

// MCP server attack chain: tool -> server -> credential -> service
MATCH path = (t:AITool)-[:UsesMCPServer]->(m:MCPServer)-[:RequiresCredential]->(c:AICredential)-[:Authenticates]->(s:AIService)
RETURN path

See BLOODHOUND_GUIDE.md located Here for the full walkthrough and cypher_queries.cy for all 29 pre-built queries.

All at once

python3 -m aihound -v --html-file report.html --json-file report.json --bloodhound bloodhound.json

Step 5: Scan Specific Tools

List what's available:

python3 -m aihound --list-tools

Available scanners:
  amazon-q             Amazon Q / AWS                 Applicable: yes
  chatgpt              ChatGPT Desktop                Applicable: yes
  claude-code          Claude Code CLI                Applicable: yes
  claude-desktop       Claude Desktop                 Applicable: yes
  cline                Cline (VS Code)                Applicable: yes
  continue-dev         Continue.dev                   Applicable: yes
  cursor               Cursor IDE                     Applicable: yes
  envvars              Environment Variables          Applicable: yes
  gemini               Gemini CLI / GCloud            Applicable: yes
  github-copilot       GitHub Copilot                 Applicable: yes
  windsurf             Windsurf                       Applicable: yes

Scan only specific tools by slug:

python3 -m aihound --tools claude-code claude-desktop envvars

What Does Each Risk Level Mean?

WSL Users

If you're running on WSL, AIHound automatically detects it and scans both:

• Linux-native paths (~/.claude/, ~/.aws/, etc.)
• Windows paths via /mnt/c/Users/<you>/AppData/...

This often reveals credentials in Windows app data that have overly permissive permissions (e.g., 0777) when viewed from WSL.

Common Findings & What They Mean

"oauth_access_token" / "oauth_refresh_token" — Claude Code

Claude Code stores OAuth tokens in ~/.claude/.credentials.json. The access token is short-lived (hours), but the refresh token is long-lived and can be used to generate new access tokens.

"mcp_env:ADO_MCP_AUTH_TOKEN" — MCP Servers

MCP server configurations often embed auth tokens directly in JSON config files. If you see inline secrets here, consider using environment variable references (${VAR_NAME}) instead.

"api_key (anthropic)" — Continue.dev

Continue.dev stores API keys in plaintext in ~/.continue/config.json. Use the ${ENV_VAR} syntax in the config to avoid this.

AWS credentials

~/.aws/credentials contains long-lived access keys. Consider using SSO/IAM Identity Center instead of static keys.

Next Steps

• Review findings and fix any CRITICAL/HIGH issues
• Generate an HTML report for your team: python3 -m aihound --html-file report.html
• Export to BloodHound for attack path visualization: python3 -m aihound --bloodhound bloodhound.json
• See BLOODHOUND_GUIDE.md for the full BloodHound walkthrough
• Check the full Documentation Here for watch mode, MCP server mode, and advanced usage