NuGuard is an open source AI application safety & security toolkit. It generates an AI Software Bill of Materials (AI-SBOM) for your agentic application, statically analyzes it for structural risk in the AI Stack and the software infrastructure. It then red-teams a sandboxed instance with a catalog of 100+ adversarial scenarios — prompt injection, tool abuse, data exfiltration, and more — so you find the issues before an attacker does. An automated judge evaluates the findings based on their impact and provides actionable remediation guidance.
What It Does
See It In Action
A real scan of a live fintech agent — Pinnacle Bank Assistant — walking through all five NuGuard stages: AI-SBOM, Cognitive Policy, Static Analysis, Behavior, and Red-Team. No mocks, no slides — real findings, including a live transcript of the agent leaking another customer's flagged fraud transactions on a routine question. → Open the interactive demo — scroll through the full walkthrough yourself.
Framework Coverage
NuGuard uses framework-aware adapters and structured configuration parsers to map the AI application stack into one evidence-backed AI-SBOM.
See the Supported Technologies documentation for the detailed language, framework, SDK, platform, Kubernetes, and datastore matrix.
Comparison
Getting Started
Install, then generate an SBOM, statically analyze it, and red-team a live target:
pip install nuguard
nuguard init --target <your-app-url>
nuguard sbom generate --source <path-to-your-app> --output app.sbom.json
nuguard analyze --sbom app.sbom.json --format markdown
nuguard redteam --config nuguard.yaml --format markdown --output reports/redteam.md
🚀 Ready to run NuGuard?
Installation, the full CLI surface, and the configuration reference — all in one guide.
🤖 Using Claude Code?
Install the NuGuard plugin and run SBOM, analysis, behavior, and red-team scans directly from Claude Code or Claude Desktop.
🔧 Using Smithery?
NuGuard is available as a Smithery Skill — install it once and use it from any MCP-compatible client (Claude Desktop, Claude Code, Cursor, and more).
# Option A — pip (recommended)
pip install "nuguard[mcp]"
python -m nuguard.mcp # starts the MCP server over stdio
# Option B — uvx (no install needed)
uvx --from "nuguard[mcp]" nuguard-mcp
Or search for nuguard on smithery.ai and click Install — Smithery configures your MCP client automatically, prompting for your LITELLM_API_KEY and other options.
🎯 Need to pentest a live target?
Nuclei-backed, bounded conventional pentest — explicit authorization required.
Hosted Version
Running NuGuard at organizational scale? The managed SaaS adds what a CISO or VP Engineering needs on top of everything in this repo — no infra to stand up or maintain.
NuGuard.ai — Managed SaaS for Security & Engineering Leaders
- 🔐 RBAC — role-based access across teams and business units
- 📊 Executive dashboards — risk posture and risk trends, board-ready
- 📋 Audit-ready reports — compliance-mapped to OWASP & MITRE ATLAS plus support for EU AI Act, NIST, etc.
- 🔗 Enterprise integrations — ServiceNow AI Control Tower, AWS Security Hub, and more
- 🛟 Managed support — dedicated onboarding and SLAs
Free trial available — no credit card required.
Contributing
🤝 Want to contribute?
Dev setup, running tests and lint, and the pull request process are covered in the Contributing guide.
Release publication is managed by repository maintainers. See Governance for ownership and the release runbook for the maintained procedure.
Repo Notes
- The repository currently contains example applications under
tests/apps/ - LLM-assisted features depend on provider credentials being available via environment variables
FAQ
I have some questions, how do I reach out to folks behind this repo? You can contact us at [email protected] For bug reporting, use the issues page on GitHub.
Do I need a live app to do a security assessment?
No. nuguard sbom + nuguard analyze find structural and supply-chain risks based on the SBOM (no source code or running application required).
nuguard behavior and nuguard redteam need a running target typically in a sandbox.
Which LLM providers are supported for LLM-assisted features?
Configured via the llm section of nuguard.yaml; provider credentials are read from environment variables. Lite LLM is used to abstract any llm provider.
For redteam scenarios, an LLM with adversarial content tolerance is required to generate attack payloads effectively. The SaaS version of nuguard provides a pre-configured LLM for this purpose, ensuring consistent and reliable red-team assessments.
What if I don't want to run all redteam scenarios?
Filter by profile (CI, Standard, Full), and configure destructive or non-destructive scenarios. Alternatively, set enabled: false per scenario in a catalog exported with nuguard redteam catalog-export.