Data sources

cudabom is grounded in first-party, machine-readable NVIDIA data. This document records the sources cudabom uses, what each is authoritative for, and how each feeds the tool. The guiding principle: cudabom does not invent fingerprints or advisories, so every claim traces back to one of these sources.

Two rules apply across all of them:

Never commit NVIDIA binaries. Archives are fetched into gitignored scratch (corpus/), verified against NVIDIA’s published SHA-256, inspected, and only the derived facts (hashes, build-ids, SONAME stems) are committed.
Scanning is offline by default. Network access is confined to explicit, developer/CI-time refresh steps (cargo xtask corpus …, cudabom db update). A normal cudabom scan needs no network.

Source hierarchy

Source	Authoritative for	cudabom use
CUDA redistributable JSON manifests	Canonical component/release inventory, versions, platforms, archive paths, SHA-256, sizes	Fingerprint corpus generation; release/component discovery; provenance
Jetson (L4T/JetPack) APT repository	Tegra aarch64 CUDA package inventory, versions, pool paths, SHA-256	Jetson fingerprint corpus (synthesized redist manifests)
CUDA package repositories (APT/deb, rpm)	OS package inventory and historical versions	OS package/version mapping (future)
NVIDIA Repo Channels / CDN (`releases.json`)	Content-addressed artifact acquisition	Future-proof download resolution (investigate)
NVIDIA Product Security (CSAF/CVE)	Security advisories	Advisory correlation / VEX (`cudabom db update`)
NGC catalog + HTTP APIs	Container metadata, declared SBOM/VEX, scan results	Real-world evaluation; declared-vs-discovered reconciliation (`cudabom reconcile`)
CUDA binary utilities / PTX ISA / programming guide docs	Format and semantics reference	Parser/reference validation, not a runtime dependency

CUDA redistributable manifests (canonical)

Base: https://developer.download.nvidia.com/compute/cuda/redist/ Per-release: .../redist/redistrib_<VERSION>.json

Each manifest is a versioned, hash-addressable catalog (not a moving “stream”): it enumerates NVIDIA components (e.g. cuda_cudart, libcublas, libcufft), their versions, per-platform archives, and each archive’s SHA-256 and size. This is the canonical source for building the fingerprint corpus. The derivation chain cudabom relies on is:

NVIDIA manifest  ->  NVIDIA archive  ->  observed binary facts  ->  cudabom fingerprint
   (says shipped)     (SHA-256 verified)   (build-id, SONAME, hash)   (committed, derived)

The cargo xtask corpus lock / corpus fetch / fingerprints build flow implements exactly this (see fingerprints/README.md). NVIDIA’s own build-system-archive-import-examples consumes redistrib_<version>.json the same way (resolve components, download, validate SHA-256, extract), which confirms the model.

Jetson (L4T/JetPack) APT repository

Base: https://repo.download.nvidia.com/jetson/common/ Per-release index: dists/<release>/main/binary-arm64/Packages

Jetson modules run Linux (L4T/JetPack), and their CUDA stack ships as Debian packages rather than the .tar.xz redistributables. The Tegra aarch64 binaries have distinct build-ids and inner-.so hashes from the generic linux-sbsa build, so without this source a scan of a real Jetson library matches only a structural family, never an exact version.

The signed APT Packages index lists, for every CUDA library package, its exact Version, pool Filename, and SHA256: the same facts a redist manifest carries. cargo xtask jetson discover reads it and synthesizes a redist-shaped redistrib_<release>.json (one archive per package, keyed under the linux-aarch64-tegra platform), so the existing corpus fetch / fingerprints build flow derives Tegra fingerprints with no special-casing. The .deb payload is unpacked during the build (its inner data.tar extracted via the object crate, which reads the signed ar archive portably). Covered releases are JetPack 5 (r35.x) and JetPack 6 (r36.x); a release that serves no CUDA package is skipped at discovery.

CUDA package repositories

Base: https://developer.download.nvidia.com/compute/cuda/repos/

Standard APT/rpm metadata (InRelease, Packages, Packages.gz) enumerating published CUDA OS packages and versions per distro/arch. Useful for OS package/version inventory, secondary to the redistrib manifests for the binary fingerprint corpus.

NVIDIA Product Security (advisories)

https://github.com/NVIDIA/product-security

NVIDIA publishes machine-readable CSAF and CVE Record files alongside Markdown bulletins. cudabom ingests these into a normalized local advisory index (advisories/index.json) via cudabom db update, kept logically separate from the fingerprint database.

NGC (evaluation corpus)

https://catalog.ngc.nvidia.com/ and https://api.ngc.nvidia.com/

NGC exposes container metadata plus NVIDIA-declared SBOM/VEX and scan results. This is the basis for a high-value validation: compare what NGC declares an image contains against what cudabom discovers in the bytes, surfacing components present but undeclared. Implemented by cudabom reconcile.

Both the SBOM and VEX are CycloneDX JSON, retrieved per image tag:

SBOM: GET {api}/v2/org/{org}/repos/{repo}/images/{tag}/sbom
VEX: GET {api}/v2/org/{org}/repos/{repo}/images/{tag}/vex

These endpoints require an NGC API key (Authorization: Bearer <key>); without one they return 401. Accordingly the fetch is opt-in and key-gated (cudabom reconcile --ngc-image org/repo:tag, key via --ngc-api-key or the NGC_API_KEY environment variable) and is the only networked path; reconciling against local --sbom/--vex files is fully offline.

Reference documentation (validation only)

CUDA binary utilities, PTX ISA, the CUDA programming guide, CUPTI, and the CUDA toolkit docs are authoritative for formats and semantics (fatbin wrapper, PTX .target, cubin ELF machine, GPU architecture numbering). They validate the parsers; they are not runtime dependencies of cudabom.