The only open-source security platform with ML-powered injection detection, MCP vulnerability scanning, and supply chain auditing. 209+ attack patterns. Zero config.
The AI security landscape is broken
Six independently useful modules. Works together or standalone. No cloud required.
Scans your MCP server configurations for SSRF vectors, exposed secrets, weak auth, and overprivileged tool definitions before deployment.
Detects malicious packages, typosquatting attacks, and GPL license violations across your agent's dependency tree.
DeBERTa-v3 model + 209+ regex patterns detects injection attacks across all known categories. Runs fully offline — no data leaves your system.
Catches injections hidden in retrieved content — HTML pages, JSON API responses, Markdown files, and zero-width character attacks.
Establishes a command baseline for each agent and flags statistical anomalies — catch unexpected behavior before it escalates to an incident.
Define allow/warn/block rules in YAML. Ship with one of 3 default profiles or compose your own stackable rule set. Version-controlled alongside your code.
From zero to protected in under 2 minutes. No account required for local use.
Install via pip. Optionally add the [ml] extra for the DeBERTa-v3 model-based detection engine.
# base
pip install shieldpilot
# with ML detection
pip install shieldpilot[ml]
Point the scanner at your MCP server configuration file. ShieldPilot auto-detects format and emits structured findings.
sentinel mcp-scan config.json
Remediate flagged findings using the built-in guidance. Install the hook once to protect all future agent sessions.
sentinel hook install
ShieldPilot is the only tool purpose-built for agentic AI security that's fully open source and self-hostable.
| Feature | ShieldPilot | Lakera | Prompt Security | DIY / Manual |
|---|---|---|---|---|
| Open Source | ✓ Apache 2.0 | ✗ | ✗ | Varies |
| Self-Hosted | ✓ | ✗ | ✗ | ✓ |
| MCP Scanning | ✓ | ✗ | ✗ | ✗ |
| ML Injection Detection | ✓ DeBERTa-v3 | ✓ | ✓ | ✗ |
| Supply Chain Audit | ✓ | ✗ | ✗ | Partial |
| Offline Mode | ✓ | ✗ | ✗ | ✓ |
| Policy-as-Code | ✓ YAML | Partial | Partial | ✓ |
| Price | Free | $$$$ | $$$$ | Engineering time |
Every detection rule, every ML model weight, every policy — visible and auditable. No black-box decisions when it comes to your agent's security posture.
# Core (no dependencies)
pip install shieldpilot
# With ML model
pip install shieldpilot[ml]
# Hook install (one-time)
sentinel hook install
# Scan MCP config
sentinel mcp-scan config.json