AI, Agents & Datasets | ESL SBOMator
AI, Agents & Datasets

Security Controls for What AI and Agents Add

SBOMator expands the inventory boundary beyond normal package manifests, providing an AI/ML profile aligned to the G7 SBOM-for-AI minimum elements โ€” covering models, datasets, MCP servers, and agent skills, all analyzed locally.

The AI / Agent Supply Chain

HalluSquatting Protection

Local intelligence for hallucinated and malicious package names, optional registry validation, updateable feeds, and real-time endpoint alerts.

Full Agent-Skill Security

SkillSpector locates lexical and hidden threats; Droid judges context, separates real risk from false positives, and feeds verdicts into a learning loop (Locate-and-Judge).

G7-Aligned AI/ML-BOM

Classifies AI components and marks output against G7 SBOM-for-AI guidance while preserving missing manual fields โ€” seven clusters: metadata, system, models, datasets, infrastructure, security, KPIs.

Model Artifacts

CycloneDX machine-learning-model components with format, path, size, and SHA-256 integrity evidence (.gguf, .safetensors, .onnx, .tflite, .pt, .h5).

MCP Inventory & Integrity

Server command, arguments, transport, and package reference, with pin-status and integrity monitoring.

ML Frameworks & Agent Frameworks

Identifies AI/LLM SDKs, agent frameworks, and vector databases with AI-role metadata.

DataBOM โ€” Dataset Provenance

Fully local scanning of AI training datasets โ€” dataset content is never uploaded.

Source & Collection Coverage

Source-domain inventory and collection-timestamp coverage.

Privacy Indicators

PII detection (email, phone, IP, payment-card) and GDPR Article 9 special-category indicators.

Guideline Checks

EDPB web-scraping guideline checks and a completeness score.

Local Dataset Readers

JSONL, CSV, and text readers with configurable record limits.

Standards-Based Output

CycloneDX 1.6 ML-BOM with data components, plus an HTML provenance report.

Evidence-Rich Findings

Secrets and provenance-risk findings surfaced as evidence.

Controlled Code Provenance

Similarity evidence without blind disclosure โ€” reduces copied-code and licensing risk without presenting a no-match result as legal clearance.

1

Git Scope

New, modified, staged, unstaged, and untracked lines by default.

2

Privacy Gate

Comments, secrets, credentials, sensitive paths, and high-entropy values removed.

3

Consent

External search requires an explicit operator decision.

4

Verify

Candidates fetched, compared locally, and recorded as evidence.

Four deployment modes. PUBLIC, PRIVATE, HYBRID, and LOCAL (local fingerprint database; no HTTP).
Reporting discipline. Voluntary G7 guidance โ€” not certification. Results are candidate provenance matches, not proof of AI origin, copyright infringement, or license incompatibility; no-match results never claim originality. Dataset content is not uploaded. Sourcegraph is an optional search backend; no partnership is implied.

Bring Your AI Supply Chain Into the Inventory

See how SBOMator maps AI components, agent capabilities, datasets, and provenance evidence without sending sensitive content away.