AI BOM & Supply Chain Transparency

The Supply Chain Record for Artificial Intelligence.

Statically scan ML models to generate secure AI Bill of Materials (AI BOMs). Detect malware, license risks, and model drift in PyTorch, Safetensors, and GGUF. New: the hosted platform turns every CI scan into a continuous inventory across your repos.

Option A: Python Package
pip install aisbom-cli 📋
Option B: Zero-Install (try first)
pipx run --spec aisbom-cli aisbom scan hf://... 📋
Option C: Air-Gapped Systems
Download Standalone Binary No Python required. Runs on bare metal.
⬇️
Explore the Hosted Platform or Read the Docs
user@dev-machine:~/project-alpha
28K+ PyPI downloads
3,400+ models scanned
Every PR scanned in CI via the GitHub Action
// LIVE SCAN

Here's a live scan of a model you've heard of.

Real output from aisbom scan hf://google-bert/bert-base-uncased. No download required — we inspect headers over HTTP.

Filename Framework Security Risk
Loading SBOM...
Scan a model now →

Why "Generic" Security Tools Miss the Mark

AI Models are binary artifacts, not just text files. Traditional scanners are blind to them.

Standard SBOM

Surface Level Scan

Only scans text manifests. Misses the actual compiled model risks.

📦 model.pt (1.2GB) ⚠ Ignored (Binary)
📦 dataset.pkl ⚠ Ignored (Binary)
AI_SBOM Deep Scan

Artifact Introspection

Decompiles tensors and inspects serialized data structures for risks.

📦 model.pt (1.2GB) ⚠ MALWARE DETECTED
↳ Found: embedded os.system() call
📦 dataset.pkl ⚠ LICENSE CONFLICT
PyTorch
HuggingFace Safetensors
GGUF / llama.cpp
Scikit-Learn (Pickle)
CycloneDX
SPDX
PyTorch
HuggingFace Safetensors
GGUF / llama.cpp
Scikit-Learn (Pickle)

Infrastructure, Not Paperwork.

Spreadsheets don't scale. We integrate directly into your training pipeline to create a dynamic, immutable record of truth.

01. Remote Streaming

Audit models on Hugging Face (hf://) without downloading terabytes of weights. We use HTTP Range requests to inspect headers and signatures over the wire.

02. Drift Detection

Stop silent changes. Run aisbom diff production.json candidate.json to detect drift in model hashes, licenses, or dependencies between commits.

03. Strict Mode Policy

Move beyond blocklists. Use --strict to enforce an allowlist-only policy, flagging any unknown import or library as a CRITICAL risk.

Ready to start?

Three ways in, all free.

Scan a model now

Try AISBOM live on a popular Hugging Face model. No download, no signup.

Scan a model now →

Connect your CI

Every PR scan builds a continuous inventory on the hosted dashboard. Free tier included.

Explore the platform →

Install the CLI

Add deep model introspection to your CI/CD pipeline in 30 seconds.

pip install aisbom-cli 📋

Frequently Asked Questions

What is an AI BOM (AI Bill of Materials)?
An AI BOM (AI Bill of Materials) is a structured inventory detailing the components, datasets, licenses, and architecture of an artificial intelligence system. It ensures supply chain security, integrity, and compliance.
How does AIsbom generate an AI BOM?
AIsbom statically disassembles Pickle bytecode and parses SafeTensors and GGUF binary headers over HTTP or locally. It inspects serialized data structures to map model dependencies, frameworks, licenses, and potential malware risks without executing any model weights.
What formats does AIsbom support for AI BOMs?
AIsbom supports industry-standard machine-readable formats including OWASP CycloneDX (v1.5 and v1.6) and SPDX (v2.3) in JSON format. This allows seamless integration into existing CI/CD pipelines and enterprise security dashboards.
Is there a hosted version of AIsbom?
Yes — the hosted platform at app.aisbom.io turns every CI scan into a continuous inventory across your repos: scan history on each PR, drift tracking, and shareable executive reports. The CLI and GitHub Action stay free and fully local; the dashboard upload is optional and opt-in, with a free tier included.
Why do traditional SBOM tools miss AI security risks?
Traditional Software Bill of Materials (SBOM) tools only inspect text-based dependency manifests. They are blind to binary model files (like .pt, .safetensors, and .gguf) which can harbor severe remote code execution (RCE) vulnerabilities and compliance-violating licenses.