v0.1.22

sbom-tools v0.1.22

Highlights

The AI BOM release. sbom-tools now treats AI systems as a first-class domain across the whole pipeline: it parses CycloneDX ML-BOMs and SPDX 3.0 AI/Dataset profiles, scores AI-readiness (AI-001…AI-010), machine-checks the EU AI Act (Annex IV) and the G7/BSI "SBOM for AI — Minimum Elements" standards, diffs model and dataset changes semantically, verifies model-weight integrity, enriches Hugging Face components, and renders AI-BOMs as a dedicated profile in the TUI (Models / Datasets / AI-Readiness tabs).

Alongside the AI work this release lands a large wave of correctness and security fixes — including a hostile-SBOM stack-overflow, empty OSV severity data, and broken CycloneDX XML parsing — a new convert command (CycloneDX ↔ SPDX), KEV/EPSS vulnerability enrichment with an offline/air-gapped mode, and broad CLI/TUI alignment.

Please read Upgrade notes below — this release changes several user-visible behaviors (scoring engine 2.1, stricter license-policy gating, logs moved to stderr, official EPSS endpoint).

What's New

AI / ML Bill of Materials

• AI-readiness scoring profile. A dedicated quality --profile ai-readiness (and TUI AI-BOM profile) grades ten AI-specific checks (AI-001…AI-010): model-card, architecture, training datasets, quantitative analysis, fairness, ethical considerations, use-cases, limitations, energy, and model-weight hash integrity. (#205, #239, #246)
• CycloneDX ML-BOM + SPDX 3.0 AI/Dataset parsing. Spec-compliant CycloneDX modelCard parsing (typed fairness/quantitative-analysis/considerations) and SPDX 3.0 AI + Dataset profile metadata including the trainedOn relationship. (#204, #208, #239)
• Semantic ML model & dataset diffing. Model and dataset changes are diffed field-by-field — quantization, architecture, training-dataset add/remove, dataset sensitivity — with high diff-costs on provenance-loss and PII-escalation signals, instead of opaque blobs. (#244)
• Model-weight integrity & Hugging Face enrichment. New AI-010 weight-hash check, a verify model-weights --model-dir subcommand (direct + HF cache layouts), and a Hugging Face enrichment source that maps siblings[].lfs.sha256 → component hashes, pipeline_tag → task, and routes pkg:huggingface models into the vulnerability stack. (#246, #253)
• SBOM-AIBOM-* SARIF rule family for AI-readiness findings, and TUI rendering of ML model / dataset metadata in the component-detail panel. (#207, #206)

Compliance

• EU AI Act (Annex IV) readiness compliance level with SBOM-AIACT-* SARIF rules, N/A-gated for non-AI SBOMs and severity-escalated by the CRA high-risk-AI flag. (#245)
• G7 / BSI "SBOM for AI — Minimum Elements" readiness — a new validate --standard bsi-ai profile that scores an AI-BOM element-by-element across the seven BSI clusters (Metadata, System-Level, Models, Datasets, Infrastructure, Security) with SBOM-BSIAI-* rules. (#255)
• Compliance rule registry. Every violation now carries a stable rule_id; SARIF rule IDs and standard references come from a single registry instead of fragile message-text matching. (#231, #233)

Cross-format conversion

• convert command. Re-emit any parsed SBOM to CycloneDX 1.7 or SPDX 2.3 with a fidelity report listing synthesized/dropped fields, built on an opt-in source-field preservation slot. (#241, #243, #247)

Vulnerability enrichment

• CISA KEV and EPSS enrichment. New --kev / --epss flags surface Known-Exploited-Vulnerabilities flags and EPSS exploit-probability on vulnerabilities, with --fail-on-kev gating, KEV/EPSS columns in reports, and a watch-mode "entered KEV" alert. (#235, #237)
• Offline / air-gapped mode + cache management. A global --offline (and SBOM_TOOLS_OFFLINE) serves enrichment purely from cache, plus a new cache status / warm / clear / export / import subcommand for sneakernet transfer. (#238)
• Unified enrichment platform. All sources (OSV, EOL, VEX, KEV, staleness, EPSS, Hugging Face) now share one EnrichmentSource trait, an atomic-write cache with schema versioning, and a single retry/User-Agent HTTP client. (#230)

CLI & TUI

• Global --config file is now honored by every command (it was previously read only by config show), with CLI-flag > file > default precedence. (#227)
• Stdin input (- as a path) for diff/quality/validate/query/ vex, enabling syft … | sbom-tools quality - pipelines. (#224)
• Typed exit codes & NDJSON output. Per-command exit codes, clap value-enums (typos fail at parse), and a wired -o ndjson format. (#226)
• Document-metadata diffing. diff now reports author, tool, timestamp, spec-version, signature and serial-number changes. (#254)
• AI-BOM is first-class in the TUI — detected as its own profile with AI-readiness scoring and dedicated Models / Datasets / AI-Readiness tabs, plus EPSS/KEV badges, a KEV filter, richer component detail, and a profile-aware help overlay. (#258, #257, #259, #261, #260, #223)

Bug Fixes

• Hostile-SBOM stack-overflow fixed. Cycle detection is now iterative (Tarjan SCC); a deeply-nested dependency chain previously aborted the process (and any FFI host) via stack overflow. (#215)
• OSV enrichment now carries severity / CVSS / description. Results from /v1/querybatch (id-only) are hydrated via /v1/vulns/{id}; previously enriched vulnerabilities had empty severity and were cached empty for 24h. (#219)
• CycloneDX XML parsing fixed for spec-conformant documents (components, dependencies, vulnerabilities, license choice, xmlns spec version). (#214)
• License policy correctness. fail_on_conflict now actually fails, concluded-only licenses are evaluated against the deny list, and deny-patterns match per-operand in SPDX expressions (deny "GPL-*" now catches MIT OR GPL-3.0-only). (#212)
• Per-component license changes are populated in diffs (the field was always empty), and the Licenses TUI tab no longer falsely reports "no changes". (#211, #259)
• Watch mode no longer fires false "resolved vulnerability" alerts on file changes (re-parsed SBOMs are now enriched before diffing). (#221)
• Incremental diff cache no longer splices sections from an unrelated cached pair, and diff errors propagate instead of becoming "no changes". (#217)
• Deterministic diffs. Fixed-seed MinHash and stable solver ordering make diff output reproducible above the LSH threshold (CI --fail-on-* gating). (#218)
• Machine output is parseable again. Logs are written to stderr so -o json / -o sarif / -o ndjson on stdout is valid when piped or redirected (previously two INFO lines preceded the JSON, breaking SARIF upload and jq). (#256)
• Multi-SBOM commands honor their flags. diff-multi/timeline/ matrix now apply --graph-*/filter/rules and can reach exit code 4, and reject unsupported -o values. (#225)
• Output hardening. CSV formula-injection guard, terminal control-char sanitization, and a TUI panic hook that restores the terminal. (#213)
• Enrichment entry-point consistency. enrich/watch/query now route through the unified orchestrator so --kev/--epss/ --huggingface take effect everywhere, query --offline no longer makes network calls, and the EPSS default endpoint is the official FIRST host with a response-size bound. (#249, #248, #247)
• FFI binding parity. The AiReadiness profile is exposed in the Go / Swift wrappers with a CI drift check, and the sbom-tools-ffi crate's version/cargo publish blockers are fixed. (#216, #234)

Performance

• Sparse assignment matching. The dense O(n²) Hungarian path (≈200 MB / effective hang on cross-format diffs) is replaced by a sparse solver over the candidate edge list. (#232)
• In-place enrichment mutation removes a 2× component-map copy per pass, and report-stage compliance is computed lazily per reporter. (#236)
• TUI dependency tab caches its graph and uses an iterative longest-path, eliminating a per-frame rebuild and an exponential depth calculation. (#223)

Internal & Quality

• Compliance engine split into a StandardChecker per-standard module layout behind a registry, and the enrichment platform unified behind one trait. (#233, #230)
• cli / tui feature gates so the FFI staticlib and fuzz targets no longer link ratatui/clap/rustls (FFI dependency tree 281 → 104 crates). (#228)
• Test infrastructure. A ratatui TestBackend snapshot harness for the TUI, golden snapshots for all report formats (plus streaming-JSON validity fixes), graph-shaped hot-path benches, and diff/scoring property tests. (#229, #240, #242, #218)

Infrastructure

• Dependency bumps: ratatui 0.30.0 → 0.30.1, chrono 0.4.44 → 0.4.45, regex 1.12.3 → 1.12.4, reqwest 0.13.3 → 0.13.4, serde_json 1.0.149 → 1.0.150, libfuzzer-sys 0.4.12 → 0.4.13, plus dagger-SDK-only bumps. New runtime dependency: flate2 (gzip, enrichment feature) for the EPSS bulk feed. (#251, #252, #195, #196, #222, #248)
• GitHub Actions bumps: codeql-action → 4.36.1, cargo-deny-action → 2.0.20, actions/checkout → 6.0.3; SHA-pin hygiene + expanded Dependabot coverage; fuzz CI now installs cargo-fuzz with the nightly toolchain. (#193, #194, #199, #200, #202)
• Total tests: 1560 (1014 lib + 546 integration), up from 1214 in v0.1.21.
• 0 clippy warnings (default + all-features) on Rust 1.88; 0 production unwrap().

Upgrade notes

These behaviors changed in this release:

• Scoring engine 2.0 → 2.1. Cycle detection now counts strongly-connected components (actual cycles) rather than DFS back-edges, so quality scores for SBOMs with cyclic dependencies may change. (#215)
• semantic_score now counts per-component license transitions, so diffs that change a component's license score slightly lower than before. (#211)
• License-policy gating is stricter. license-check now fails on license conflicts and concluded-only denied licenses that previously passed silently — CI gates may newly fail (correctly). (#212)
• Logs are on stderr. Anything parsing tool stdout for log lines must now read stderr; machine-readable report output on stdout is clean. (#256)
• EPSS default endpoint moved to the official FIRST host epss.empiricalsecurity.com (gzip). Override with --epss-url / SBOM_TOOLS_EPSS_URL. (#248)
• Library API: IncrementalDiffEngine::diff and MultiDiffEngine::{diff_multi,timeline,matrix} now return Result. (#217)

Acknowledgments

The AI-BOM direction in this release was driven by @MChorfa (Mohamed Chorfa), who proposed the original ML-BOM and AI-readiness work and filed the issues this release implements: AI-BOM model-card scoring (#184), SBOM-AIBOM SARIF rules (#185), SPDX 3.0 AI-profile parsing (#186), and TUI ML/dataset display (#187). Thanks also to @jkowalleck (Jan Kowalleck) for the CycloneDX Tool Center invitation (#176) behind the new submission guide (#177), and continued thanks to @cmyank0 and @VincentR-OCD whose issue reports keep sbom-tools' diff and enrichment output honest.

This release also builds on the public security and standards work behind the G7/BSI "SBOM for AI" minimum elements, the EU AI Act, CISA KEV, and FIRST EPSS.

---

Install: cargo install sbom-tools
Homebrew: brew install sbom-tool/tap/sbom-tools
Crate: https://crates.io/crates/sbom-tools
Full changelog: v0.1.21...v0.1.22