Skip to content
Back to skills

Ultrasafe Supply Chain Auditor

ASecurity

Pre-release simulated penetration testing from the dependency / SBOM / typosquatting / signing-chain perspective. Invoke as Agent 3 of the Ultrasafe 8-agent fan-out (Ultrasafe.md §15.3) whenever the iteration axis-set includes `usf-supply-chain`, or whenever a release candidate touches dependency manifests (package.json / pyproject.toml / Cargo.toml / go.mod / requirements*.txt / lockfiles / SBOM artifacts / signed-attestation chains). Emits `ULTRASAFE_FINDING` A2A intents (Constellation §13....

  • 8 stars
  • 0 votes
  • 0 copies
  • 3 views
  • Added September 29, 2026
ai-agentspythonrustgobashnodetestinggitsecuritydocumentation

Works with

  • cursor
  • cli

Security analysis

A100/100

Scanned September 29, 2026

npx -y skills add SoliEstre/EstreGenesis --skill ultrasafe-supply-chain-auditor --agent claude-code

Installs into .claude/skills of the current project.

Are you the author of Ultrasafe Supply Chain Auditor?

Add the live security badge to your README. It updates with every re-scan.

Security grade badge for Ultrasafe Supply Chain Auditor
[![Security: A — Skills Directory](https://www.skillsdirectory.com/api/skills/soliestre-ultrasafe-supply-chain-auditor/badge)](https://www.skillsdirectory.com/skills/soliestre-ultrasafe-supply-chain-auditor)

More formats (shields.io, HTML) on the badges page. Keep it an A: scan every change in CI with Pro.

Download with Pro
SKILL.md
---
name: ultrasafe-supply-chain-auditor
version: 0.2.0
description: Pre-release simulated penetration testing from the dependency / SBOM / typosquatting / signing-chain perspective. Invoke as Agent 3 of the Ultrasafe 8-agent fan-out (Ultrasafe.md §15.3) whenever the iteration axis-set includes `usf-supply-chain`, or whenever a release candidate touches dependency manifests (package.json / pyproject.toml / Cargo.toml / go.mod / requirements*.txt / lockfiles / SBOM artifacts / signed-attestation chains). Emits `ULTRASAFE_FINDING` A2A intents (Constellation §13.16, §18.1) with PURL canonical ids, OSV CVE matches, SLSA provenance verdicts, cosign+Rekor attestation chain status, and maintainer-anomaly flags. v0.2.x = advisory mode (report-only, never auto-blocks publish); v0.3+ = blocking gate for deterministic signals only. Maintainer-anomaly findings are always human-gated (Ultrasafe §2.1.3 CT1) regardless of mode.
---

# Supply Chain Auditor — Ultrasafe Attacker Skill

> **Role**: Pre-release simulated penetration testing from the dependency vulnerability / SBOM mismatch / typosquatting / signing-chain perspective.
> **Tone**: dependency-graph-aware — every finding is anchored to a node in the dependency DAG, with the transitive path made explicit (root → … → vulnerable node) and the build/publish lane (npm, PyPI, crates.io, Go module proxy, container registry) named.
> **Output**: `ULTRASAFE_FINDING` A2A intents (Constellation §13.16, runtime wire = Ultrasafe.md §18.1) — advisory only in v0.2.x.
> **Position**: Agent 3 of the 8-agent Ultrasafe fan-out (Ultrasafe.md §15.3).

## §1 When to invoke

Fire this skill in any of the following situations:

1. **Iteration axis-set includes `usf-supply-chain`** — orchestrator (Ultrasafe.md §15, §9.9) dispatches this skill automatically as part of the parallel fan-out, regardless of tier. SCS 5-way coverage (build / maintainer / typo / transitive / reproducibility) is mandatory whenever this axis is active.
2. **Pre-release trigger touches a dependency manifest** — the PreToolUse hook (`hooks/ultrasafe-trigger.cjs`, Ultrasafe.md §17.1) detects an imminent `npm publish` / `pip upload` / `cargo publish` / `git push --tags` and any of `package.json` / `package-lock.json` / `pnpm-lock.yaml` / `pyproject.toml` / `poetry.lock` / `requirements*.txt` / `Cargo.toml` / `Cargo.lock` / `go.mod` / `go.sum` / `Gemfile.lock` / `composer.lock` / `Pipfile.lock` / SBOM file (CycloneDX `bom.json`, SPDX `*.spdx.json`) has changed since the last clean iteration baseline.
3. **A new transitive dependency appears** — diff between current commit and `last_clean_baseline_sha` introduces a new node in the resolved dependency DAG (Ultrasafe.md §6.5 — promotes Tier 1 patch to Tier 2).
4. **External `SECURITY_DISCLOSURE_INTAKE` references a CVE in a current dependency** — Ultrasafe §18.4 inbound disclosure intent names a CVE / GHSA / OSV id that matches a PURL in the current SBOM.
5. **Signing-chain artifact added or rotated** — cosign keyless certificate, Rekor entry, SLSA provenance attestation, or in-toto layout changes since the last iteration; re-verify the entire chain before allowing the new artifact to participate in clean-signal evaluation.

**Skip** if:

- All five SCS 5-way signals were green in the previous iteration AND no dependency manifest / lockfile / SBOM / signing artifact has changed (delta-only Tier 1 patch lane, Ultrasafe.md §6.3 — but only the manifest-static portion of `usf-supply-chain` runs).
- The iteration is a documentation-only commit (no `.toml` / `.json` / `.lock` / `.mod` / `.sum` / SBOM touched).

## §2 Input

Structured input received from the orchestrator (Ultrasafe.md §15.3):

```json
{
  "target_commit_sha": "string (40-char hex)",
  "catalog_versions": {
    "SLSA": "v1.0",
    "NIST_SSDF": "SP 800-218 v1.1",
    "OSV": "<query date ISO-8601>",
    "Sigstore": "<rekor public-good instance version>",
    "CycloneDX": "1.5 | 1.6",
    "SPDX": "2.3"
  },
  "iteration": 1,
  "prior_findings_set": ["finding_id_1", "finding_id_2", "..."],
  "sbom_path": "path/to/bom.json",
  "lockfile_paths": ["package-lock.json", "poetry.lock", "..."],
  "attestation_paths": ["*.sig", "*.crt", "*.intoto.jsonl", "..."],
  "threat_model_ref": "optional — Ultrasafe.md §11 threat model snapshot id",
  "tier": "1 | 2 | 3",
  "axis_set": ["usf-supply-chain", "..."]
}
```

The `prior_findings_set` lets you detect **regression** (a previously-resolved finding that has reappeared — Ultrasafe.md §15.0 termination condition 1) and avoid duplicate-emit storms.

## §3 Methodology — SCS 5-way attacker probe matrix

Walk the dependency graph as an adversary would. For each of the five vectors, run the structured probe list below and collect evidence. **Tools available**: `Read`, `Grep`, `Bash` (read-only — `cosign verify`, `osv-scanner`, `npm audit --json`, `pip-audit --format=json`, `cargo audit --json`, `govulncheck -json`, `syft`, `grype`, `slsa-verifier`, `rekor-cli search`; **no outbound endpoint calls** — operate against local OSV cache, local Rekor mirror snapshot, and local SBOM only).

### §3.1 Vector 1 — Build-time compromise (CI / runner / pipeline tampering)

Probes:
- Verify SLSA Build Level for every artifact in the SBOM. Anything below the declared minimum (Tier 1 → L1, Tier 2 → L2, Tier 3 → L3) is a finding.
- Check the GitHub Actions workflow files (`.github/workflows/*.yml`) for unpinned third-party actions (any `uses: <org>/<repo>@<tag>` where `<tag>` is not a 40-char commit SHA).
- Diff CI config against the previous iteration's clean baseline; any new `secrets:` reference or new `permissions:` widening is a finding.
- Re-run the provenance generator against the artifact and compare digest; mismatch is a build-time tamper finding.
- Look for `npm postinstall` / `pip setup.py` / `cargo build.rs` hooks added by transitive dependencies — each one is a probe target.

Evidence: SLSA verifier output (`slsa-verifier verify-artifact ...`), workflow file diffs, postinstall hook listings.

### §3.2 Vector 2 — Maintainer anomaly (account takeover, sudden ownership change, sock-puppet review)

Probes:
- For each direct dependency, check OSV / GitHub Security Advisory / npm advisories for recent maintainer compromise reports.
- Inspect `npm view <pkg> maintainers` / PyPI `users` / crates.io `owners` and compare to the previous iteration's snapshot — any new owner in the last 30 days is flagged.
- Detect single-maintainer dependencies (bus-factor = 1) on the critical path — not auto-block, but flag for human review.
- Detect "ownership flip then immediate version bump" pattern (the classic event-stream / ua-parser-js / pyutil playbook).

Evidence: maintainer list snapshot diff, registry metadata, ownership transfer timestamps.

**Auto-block prohibition** (Ultrasafe.md §2.1.3 cross-axis Contradiction CT1): maintainer anomaly is **never** auto-blocking — always emits with `value.human_gate_required: true` and routes to the human-review lane on the live board.

### §3.3 Vector 3 — Typosquatting / dependency confusion / brandjacking

Probes:
- For each direct dependency, compute Levenshtein distance ≤ 2 against the top-1000 most-downloaded packages on the same registry. Distance-1 matches against a popular package are critical (e.g. `requets` vs `requests`).
- Detect homoglyph attacks (Cyrillic `а` vs Latin `a`, etc.) in package names.
- Check for **dependency confusion** — any internal-scoped package name (`@<org>/<pkg>`) that also exists on the public registry with a higher version number.
- Scan for newly-added scoped packages whose unscoped form already exists publicly.

Evidence: registry lookups (local cache only), Levenshtein distance computation, scope/version comparison tables.

### §3.4 Vector 4 — Transitive vulnerability (deep-tree CVE, EPSS-weighted)

Probes:
- Run `osv-scanner --lockfile=<each lockfile> --offline` against local OSV cache; aggregate by PURL.
- For each CVE match, record CVSS v4 base, EPSS estimate, and KEV inclusion (CISA Known Exploited Vulnerabilities — KEV is a trump signal per Ultrasafe.md §6.1).
- Compute the **transitive path** for every vulnerable node: `root_pkg → A → B → vulnerable_node`. Path-length > 3 is dependency-graph-aware context worth flagging; path-length 1 (direct) is highest priority.
- Cross-reference findings against `prior_findings_set` to detect regression (previously-resolved CVE has reappeared due to lockfile rollback or pin removal).

Evidence: OSV scanner JSON output, transitive path graph, CVSS/EPSS/KEV table.

### §3.5 Vector 5 — Reproducibility failure (build determinism, source-to-binary integrity)

Probes:
- For dependencies that publish source + binary (Python wheels, prebuilt binary npm packages, Go module zips), verify that rebuilding from the published source yields the same digest as the published binary (where reproducible-builds tooling exists).
- Verify cosign / Sigstore signatures and Rekor inclusion proofs for every signed artifact:
  - `cosign verify <artifact> --certificate-identity ... --certificate-oidc-issuer ...`
  - `rekor-cli get --uuid <uuid>` against local Rekor mirror snapshot.
- Verify in-toto attestation chains end-to-end; gaps (missing intermediate step) are findings.
- Check SBOM-to-installed-artifact match: every package in the lockfile must appear in the SBOM with matching version + integrity hash; **SBOM mismatch** is a finding (Ultrasafe.md §15.3 tone "SBOM mismatch").

Evidence: cosign verify output, Rekor inclusion proof, SBOM diff vs installed artifacts.

### §3.6 Severity computation

Use the Ultrasafe.md §6.1 priority matrix:

```
priority = max(
  CVSS_v4_base,                      // baseline
  EPSS_estimate × 10,                // exploitation likelihood weight
  KEV ? 10 : 0,                      // KEV is automatic critical
  deterministic_signal ? 9 : 0       // cosign mismatch / SLSA L0 / OSV direct CVE
)
```

Severity bucket:
- `critical` — priority ≥ 9 OR KEV match
- `high` — 7 ≤ priority < 9
- `medium` — 4 ≤ priority < 7
- `low` — 1 ≤ priority < 4
- `info` — observation only (e.g. bus-factor = 1 with no other signal)

### §3.7 Evidence labeling discipline

Every finding's `evidence` field MUST be prefixed with one of:
- `[verified]` — reproducible from local tools (osv-scanner / cosign / slsa-verifier output included).
- `[inferred]` — derived from registry metadata or maintainer listings (not independently reproducible without registry access).
- `[assumed]` — heuristic-only (e.g. typosquatting Levenshtein distance is suggestive, not confirmed malicious).

Mixing labels in a single finding is forbidden — split into multiple findings instead. This anchors the cross-axis Contradiction discipline (Ultrasafe.md §2.1).

## §4 Finding output schema

Each finding is emitted as a single `ULTRASAFE_FINDING` A2A intent (Constellation §13.16, runtime wire = Ultrasafe.md §18.1):

```json
{
  "type": "CUSTOM",
  "name": "ULTRASAFE_FINDING",
  "targetAgentId": "main",
  "value": {
    "finding_id": "us-2026-06-06-scs-001",
    "iteration": 1,
    "axis": "usf-supply-chain",
    "agent_id": "ultrasafe-supply-chain-auditor",
    "attack_pattern": {
      "stix_id": "attack-pattern--<uuid>",
      "mitre_technique": "T1195.002"
    },
    "severity": {
      "cvss": 8.4,
      "epss_estimate": 0.27,
      "ultrasafe_exploited": false,
      "asr_pct": 0,
      "ci95": [0, 0]
    },
    "diamond": {
      "adversary": "supply-chain-actor (unknown)",
      "capability": "transitive CVE exploit",
      "infrastructure": "npm registry",
      "victim": "downstream consumers"
    },
    "category": "transitive-vulnerability | sbom-mismatch | typosquatting | signing-chain-gap | maintainer-anomaly | build-tampering | reproducibility-failure | dependency-confusion",
    "purl": "pkg:npm/<name>@<version>",
    "transitive_path": ["root", "@scope/a@1.0.0", "b@2.3.4", "vulnerable@0.1.7"],
    "cve_ids": ["CVE-2026-XXXXX", "GHSA-xxxx-xxxx-xxxx"],
    "kev_match": false,
    "attestation_chain": {
      "cosign_verified": true,
      "rekor_inclusion": true,
      "slsa_level_observed": "L2",
      "slsa_level_required": "L3",
      "intoto_layout_complete": true
    },
    "maintainer_anomaly_flag": false,
    "evidence": "[verified] osv-scanner reports CVE-2026-XXXXX (CVSS 8.4, EPSS 0.27) in pkg:npm/vulnerable@0.1.7 reached via root → @scope/a@1.0.0 → b@2.3.4 → vulnerable@0.1.7; raw output excerpt: ...",
    "evidence_ref": "sarif://local/iteration_1/scs/finding_001.sarif",
    "reproduction_steps": "1. cd repo && npm ci  2. osv-scanner --lockfile=package-lock.json --offline  3. observe CVE-2026-XXXXX entry  4. npm ls vulnerable to confirm transitive path",
    "recommended_fix": "Pin @scope/a to ≥1.2.0 (resolves b to ≥3.0.0, which pins vulnerable to ≥0.2.0). Alternative: add `overrides` entry to force vulnerable@0.2.0. Verify fix by re-running osv-scanner and confirming no CVE match.",
    "redaction": "external_summary_only",
    "spotlight_wrap": true,
    "advisory": true,
    "would_block_in_v03_blocking": true,
    "human_gate_required": false,
    "external_standard_anchor": {
      "catalog_version": "OSV 2026-06-06",
      "catalog_cell_id": "CVE-2026-XXXXX",
      "coverage_pct_under_catalog": 100
    }
  }
}
```

Field discipline:
- `advisory: true` is **mandatory** in v0.2.x — every emit is report-only.
- `would_block_in_v03_blocking` is the future-blocking hint: `true` only for **deterministic signals** (cosign signature mismatch, missing SLSA provenance, direct OSV CVE match against a non-disputed PURL).
- `human_gate_required: true` is **mandatory** for any maintainer-anomaly finding (Ultrasafe.md §2.1.3 CT1) regardless of mode.
- `category` is one of the eight enumerated values listed above — no free-form categories.
- `purl` MUST be a canonical Package URL (PURL spec) — `pkg:<type>/<namespace>/<name>@<version>`.
- `transitive_path` is ordered root-to-leaf; for direct dependencies, the array has length 2 (`["root", "<pkg>@<ver>"]`).
- `cve_ids` is empty `[]` for non-CVE findings (typosquatting, SBOM mismatch, signing-chain gap with no associated CVE).

## §5 Examples

### Example 1 — Trigger via axis dispatch (Tier 2 iteration 1, transitive CVE)

Orchestrator dispatches `ultrasafe-supply-chain-auditor` for iteration 1 of a Tier 2 release candidate. Input includes `axis_set: ["usf-supply-chain", "usf-web-sast-dast", ...]` and `tier: 2`. The skill:

1. Runs `osv-scanner --lockfile=package-lock.json --offline` and detects `CVE-2026-XXXXX` (CVSS 8.4, EPSS 0.27) in transitive dependency `vulnerable@0.1.7`.
2. Computes transitive path: `root → @scope/a@1.0.0 → b@2.3.4 → vulnerable@0.1.7`.
3. Emits a single `ULTRASAFE_FINDING` with `category: "transitive-vulnerability"`, severity bucket `high` (priority = max(8.4, 2.7, 0, 9) = 9), `advisory: true`, `would_block_in_v03_blocking: true` (deterministic OSV match), `human_gate_required: false`.
4. Provides reproduction steps using only the local OSV cache and recommended fix (pin @scope/a to ≥1.2.0).

### Example 2 — Trigger via PreToolUse hook (signing-chain gap before npm publish)

User runs `npm publish` for a Tier 3 release. The PreToolUse hook (Ultrasafe.md §17.1) intercepts and dispatches the 8-agent fan-out. `ultrasafe-supply-chain-auditor` runs and:

1. Executes `cosign verify-blob --signature dist.sig --certificate dist.crt dist.tar.gz` against the local Sigstore trust root.
2. cosign succeeds but `slsa-verifier verify-artifact dist.tar.gz --provenance-path dist.intoto.jsonl --source-uri github.com/<org>/<repo>` returns "no provenance found".
3. Emits `ULTRASAFE_FINDING` with `category: "signing-chain-gap"`, severity `high`, `attestation_chain.slsa_level_observed: "L0"`, `attestation_chain.slsa_level_required: "L3"`, `advisory: true`, `would_block_in_v03_blocking: true`.
4. Recommended fix: enable `slsa-github-generator` workflow + regenerate provenance + re-attest.
5. **v0.2.x behavior**: the publish proceeds (advisory mode = report-only), the finding appears on the live board with the `ADVISORY` corner badge, and the iteration boundary records `clean_signal_4_condition_AND_gate_state.condition_3_coverage_floor: false`.

### Example 3 — Trigger via maintainer-anomaly detection (Tier 2, human-gate required)

During iteration 2 of a Tier 2 release, the maintainer listing for direct dependency `popular-utility@4.2.0` shows a new owner added 6 days ago, followed by a version bump from 4.1.x to 4.2.0 yesterday. `ultrasafe-supply-chain-auditor`:

1. Diffs current maintainer snapshot against the previous iteration's clean baseline; detects the new owner.
2. Cross-references the version-bump timing; the "ownership flip → immediate bump" pattern matches the historical event-stream / ua-parser-js / pyutil playbook.
3. Emits `ULTRASAFE_FINDING` with `category: "maintainer-anomaly"`, severity `medium` (no CVE yet, but signal is suggestive), `evidence` prefixed `[inferred]`, `maintainer_anomaly_flag: true`, **`human_gate_required: true`** (Ultrasafe.md §2.1.3 CT1), `would_block_in_v03_blocking: false` (NOT a deterministic signal — never auto-block, even in blocking mode).
4. Recommended fix: hold the dependency at 4.1.x pending human review, monitor OSV / GitHub Security Advisories for the package over the next 14 days.

### Example 4 — Trigger via inbound `SECURITY_DISCLOSURE_INTAKE` referencing a current dependency

Inbound `SECURITY_DISCLOSURE_INTAKE` intent (Ultrasafe.md §18.4) names CVE-2026-YYYYY in `crypto-lib@2.1.0`. `ultrasafe-supply-chain-auditor`:

1. Greps the SBOM for `crypto-lib@2.1.0` — found, as a transitive dependency via `tls-helper@3.0.4`.
2. Verifies the CVE applies to the version-range claimed (OSV record vs PURL).
3. Emits `ULTRASAFE_FINDING` with `category: "transitive-vulnerability"`, `cve_ids: ["CVE-2026-YYYYY"]`, severity per §3.6 priority matrix, `evidence` prefixed `[verified]` if OSV record is reachable in local cache, `[inferred]` if relying on the inbound disclosure body alone.
4. Cross-links to the inbound `SECURITY_DISCLOSURE_INTAKE` envelope via `evidence_ref` containing the inbox cursor reference.

## §6 Anti-patterns

Do NOT do any of the following:

1. **Calling outbound endpoints during scanning.** Tools must operate against local caches (OSV mirror, Sigstore trust root snapshot, registry metadata snapshot). Any outbound HTTP call from this skill is a violation of Ultrasafe.md §15.3 ("외부 endpoint 호출 금지 — 로컬 cache 만 사용"). If the local cache is stale, emit a finding about the stale cache instead of fetching.
2. **Auto-blocking on maintainer anomaly.** Maintainer anomaly is **always** human-gated (Ultrasafe.md §2.1.3 CT1). Even in v0.3+ blocking mode, this signal class never auto-blocks. Emitting `would_block_in_v03_blocking: true` on a `category: "maintainer-anomaly"` finding is a violation.
3. **Aggregating multiple CVEs into one finding.** Each CVE / PURL pair gets its own `ULTRASAFE_FINDING` emit. Bundling lets the live board's Tier 2 card surface (Ultrasafe.md §8.5) lose granularity and breaks the §13.16 outbox.jsonl audit trail.
4. **Using the word "secure" without a coverage attestation.** Methodology discipline (Ultrasafe.md §15.6 §2.1.6): never write `"this dependency is secure"`. Instead: `"passed coverage 100% under OSV catalog v2026-06-06 as of 2026-06-06T14:00Z"`. Same rule applies to supply-chain findings.
5. **Skipping the `[verified] | [inferred] | [assumed]` evidence label.** Every `evidence` field must carry one of the three prefixes. Unlabeled evidence breaks the cross-axis Contradiction discipline (Ultrasafe.md §2.1) — the Synthesizer (§15.8) cannot dedup or rank without it.
6. **Emitting a finding without `purl` for a dependency-anchored category.** Categories `transitive-vulnerability` / `sbom-mismatch` / `typosquatting` / `dependency-confusion` / `maintainer-anomaly` MUST include a canonical PURL. Findings without a PURL are not dependency-graph-aware and violate the tone contract.
7. **Mixing advisory and blocking outputs in one emit.** v0.2.x is advisory mode end-to-end. Every emit MUST set `advisory: true`. The `would_block_in_v03_blocking` field is a *hint about future behavior*, not a current state. Misusing it as a current blocking flag is a violation.
8. **Re-emitting findings from `prior_findings_set` without checking regression.** If a finding id from `prior_findings_set` reappears, that is a **regression event** (Ultrasafe.md §15.0 termination condition 1) and MUST be flagged in the finding's `evidence` as such. Silently re-emitting wastes synthesizer cycles and hides the regression signal from the iteration-boundary clean-signal AND-gate (Ultrasafe.md §18.2 condition 1).

## §7 Cross-references

- **Ultrasafe.md §15.3** — Agent 3 role specification (input / output / tools / when-to-fire / severity rubric).
- **Ultrasafe.md §2.1.3** — cross-axis Contradiction CT1 (maintainer anomaly never auto-blocks).
- **Ultrasafe.md §6.1** — CVSS v4 + EPSS + KEV priority matrix.
- **Ultrasafe.md §6.3** — Tier 1/2/3 dispatch matrix (which agents run when).
- **Ultrasafe.md §8.5** — live board Tier 2 finding-card surface.
- **Ultrasafe.md §15.0** — 8-agent fan-out termination conditions (clean-signal 4-condition AND-gate).
- **Ultrasafe.md §15.8** — Synthesizer (cross-axis dedup + severity rank + correlation).
- **Ultrasafe.md §17.1** — PreToolUse hook `ultrasafe-trigger.cjs` (publish-equivalent matchers).
- **Ultrasafe.md §18.1** — `ULTRASAFE_FINDING` A2A runtime wire spec.
- **Ultrasafe.md §18.4** — `SECURITY_DISCLOSURE_INTAKE` (inbound disclosure intake that can trigger this skill).
- **Constellation.md §13.16** — A2A intent allowlist registration + ack_tier + Spotlighting wrapper.
- **Constellation.md §13.13** — at-least-once delivery + commitment ack tier.
- **Greatpractice.md** (when published) — supply-chain auditor finding patterns are candidate tree nodes under `pre-release / dependency-discipline / signing-chain` macro branches.
- **External standards**: SLSA v1.0, NIST SSDF SP 800-218 v1.1, OSV schema, Sigstore (cosign + Rekor), CycloneDX 1.5/1.6, SPDX 2.3, MITRE ATT&CK T1195 (Supply Chain Compromise).

Attribution

Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.

Comments

Loading comments…

Newsletter

Get the weekly list of safe new skills, and the ones we pulled.