[emmanuelgjr]

Summary

This PR implements items 1–6 and 14 of the data-validation review: the places where the validation lagged current standards or silently skipped a check.

#	Change
1	OWASP Top 10 for LLM Applications 2026. The 2026 edition (published 2026-08-04, canonical repo) renumbered the list; for example, LLM03 changed from Supply Chain to Excessive Agency. Schemas now accept LLMnn:2025 and LLMnn:2026, always with the year. reference_data/owasp_top10.csv holds both editions plus a 2026→2025 crosswalk, which I used to add the 2026 equivalent to every existing mapping (111 files, 163 IDs; nothing removed).
2	OWASP Top 10 for Agentic Applications 2026. New optional owasp_agentic_top10_mapping field (ASI01:2026…ASI10:2026) in the exploit, vulnerability and agent-trace schemas. OWASP writes these IDs without a year; the suffix keeps a future edition from colliding with this one, as just happened with the LLM list.
3	CVSS v4.0. Adds cvss_v4_score and cvss_v4_vector. The qualitative rating "Informational" becomes FIRST's "None" (no entry used it). A new severity_check.py recomputes every score from its vector (v3.0/3.1/4.0) and checks that the qualitative rating matches the score's FIRST band. Data fixes found by the check, against NVD: • CVE-2024-37032: 8.8 is High, not Critical. • CVE-2025-0845: the published vector gives 6.4, not 6.5. • Vectors added for CVE-2024-34359, CVE-2025-0845 and CVE-2025-32711 (the CNA vector, which matches the stored 9.3).
4	Format enforcement. The exploit, vulnerability and risk-assessment validators built their validator without a format checker, so "format": "date" and "uri" were never checked. They now check both, and jsonschema[format-nongpl] supplies the URI checker. All current data passes.
5	JSON Schema 2020-12 for incident, rag, promptinj_testcase and crossframework, matching the other datasets. The Turkish pairs validator still passes against the migrated promptinj schema.
6	Python 3.11+, because 3.10 reaches end of life this month. CI runs the tests on 3.11, 3.12 and 3.13.
14	reference-data-refresh workflow. Runs on the 2nd of each month (and on demand) and regenerates the ATLAS and CWE tables. If anything changed, it pushes automation/reference-data-YYYY-MM and opens an issue with the compare link: this repo doesn't let workflows open PRs (can_approve_pull_request_reviews: false), and I didn't loosen that setting.

The new tests/test_standards.py fails if a schema drifts from 2020-12, if the OWASP enums stop matching the reference table, or if a dataset validator drops its format checker.

Verification

• run_all_checks.py: 7 validators pass. The shared checks find 0 errors across 591 files; the only 2 warnings are the intentional retired-ID notes from exploit_dataset: move retired ATLAS IDs to AML.T0115 sub-techniques #90.
• 122 tests pass on Python 3.12 and 3.13 (1 skipped: the contributor-maintained contrastive schema). agentdataflow's own 18 tests pass.
• actionlint is clean on both workflows.

🤖 Generated with Claude Code

[github-actions]

Dependency Review

The following issues were found:

• ✅ 0 vulnerable package(s)
• ❌ 1 package(s) with incompatible licenses
• ✅ 0 package(s) with invalid SPDX license definitions
• ✅ 0 package(s) with unknown licenses.

See the Details below.

License Issues

data_validation/requirements.txt

Package	Version	License	Issue Type
cvss	>= 3.2	LGPL-3.0 AND LGPL-3.0-only AND LGPL-3.0-or-later	Incompatible License

Allowed Licenses:

Apache-2.0, BSD-2-Clause, BSD-3-Clause, ISC, MIT, MPL-2.0, CC-BY-4.0, CC0-1.0, Unlicense

OpenSSF Scorecard

Package	Version	Score	Details
actions/actions/checkout	3d3c42e5aac5ba805825da76410c181273ba90b1	🟢 6.3	Details Check Score Reason Maintained 🟢 4 5 commit(s) and 0 issue activity found in the last 90 days -- score normalized to 4 Code-Review 🟢 10 all changesets reviewed Dangerous-Workflow 🟢 10 no dangerous workflow patterns detected Binary-Artifacts 🟢 10 no binaries found in the repo Token-Permissions ⚠️ 0 detected GitHub workflow tokens with excessive permissions CII-Best-Practices ⚠️ 0 no effort to earn an OpenSSF best practices badge detected Pinned-Dependencies 🟢 3 dependency not pinned by hash detected -- score normalized to 3 Fuzzing ⚠️ 0 project is not fuzzed License 🟢 10 license file detected Signed-Releases ⚠️ -1 no releases found Security-Policy 🟢 9 security policy file detected Packaging ⚠️ -1 packaging workflow not detected SAST 🟢 10 SAST tool is run on all commits Branch-Protection 🟢 6 branch protection is not maximal on development and all release branches
actions/actions/setup-python	5fda3b95a4ea91299a34e894583c3862153e4b97	🟢 6.6	Details Check Score Reason Maintained 🟢 10 17 commit(s) and 0 issue activity found in the last 90 days -- score normalized to 10 Binary-Artifacts 🟢 10 no binaries found in the repo Code-Review 🟢 10 all changesets reviewed Dangerous-Workflow 🟢 10 no dangerous workflow patterns detected Packaging ⚠️ -1 packaging workflow not detected Token-Permissions ⚠️ 0 detected GitHub workflow tokens with excessive permissions Pinned-Dependencies 🟢 7 dependency not pinned by hash detected -- score normalized to 7 CII-Best-Practices ⚠️ 0 no effort to earn an OpenSSF best practices badge detected Fuzzing ⚠️ 0 project is not fuzzed License 🟢 10 license file detected Signed-Releases ⚠️ -1 no releases found Security-Policy 🟢 9 security policy file detected Branch-Protection ⚠️ 0 branch protection not enabled on development/release branches SAST 🟢 9 SAST tool is not run on all commits -- score normalized to 9
pip/cvss	>= 3.2	Unknown	Unknown

Scanned Files

• .github/workflows/reference-data-refresh.yml
• data_validation/requirements.txt

[emmanuelgjr]

Dependency Review flagged the cvss package as LGPL-3.0, which the repo's licence policy doesn't allow. I left the policy alone and replaced the package with validators/cvss_score.py:

• v3.0/v3.1 use the FIRST specification formulas.
• v4.0 is a port of FIRST's BSD-2-Clause reference calculator (cvss40.js); its licence and lookup tables are included.
• The port matches the reference on 500 vectors (tests/test_cvss_score.py), and the v3 scorer reproduces the NVD/CNA scores of the dataset's CVEs.

The tooling is back to one runtime dependency (jsonschema).