After showing that vanilla Claude Sonnet 4.6 and Opus 4.8 both missed CVE-2026-27205 (Flask 2026 session fixation), we asked the obvious follow-up: does loading a comprehensive cybersecurity skills library into Claude change the result? We installed the mukul975/Anthropic-Cybersecurity-Skills pack — 817 skills, v1.3.0 — and re-ran the same supply-chain test.
Short answer: skills do not add live CVE data. They teach methodology. The gap closed only when the recommended tools were actually executed — not when the skills were loaded.
Setup
npx skills add mukul975/Anthropic-Cybersecurity-Skills
# → 817 skills installed across .claude/skills/ and .agents/skills/Target dependencies (intentionally vulnerable):
flask==2.3.0 ← CVE-2026-27205 (session fixation, CVSS 5.0) — 2026
krb5==0.7.1 ← MGASA-2026-0233 / CVE-2026-40355 (unauthenticated RCE)
gssapi==1.8.2 ← same advisory, gss_accept_sec_context() vulnerable
Werkzeug==2.3.7 ← CVE-2024-34069 (debugger RCE), CVE-2023-46136
requests==2.31.0 ← CVE-2024-35195, CVE-2026-25645Skills activated for this test:
analyzing-sbom-for-supply-chain-vulnerabilities— parse SBOM, query NVD 2.0, grype cross-validationhunting-for-supply-chain-compromise— threat hunting for compromised PyPI/npm packagesperforming-sca-dependency-scanning-with-snyk— pip-audit, snyk, OWASP Dependency-Checkdetecting-golden-ticket-attacks-in-kerberos-logs— Kerberos TGT anomaly detection
Test 1 — Skills-armed static analysis (no tool execution)
Claude with skills loaded, asked to identify all CVEs using only its built-in knowledge:
Flask 2.3.0:
✓ CVE-2023-30861 (CVSS 7.5) [2023, pre-cutoff]
✗ CVE-2026-27205 (CVSS 5.0) [2026 — NOT FOUND]
Werkzeug 2.3.7:
✓ CVE-2023-46136, CVE-2024-34069 [pre-cutoff]
✗ CVE-2026-21860, CVE-2026-27199 [2026 — NOT FOUND]
requests 2.31.0:
✓ CVE-2024-35195
✗ CVE-2026-25645 [2026 — NOT FOUND]
krb5 0.7.1 (Python bindings):
✓ CVE-2022-42898, CVE-2023-36054 [C library, pre-cutoff]
✗ CVE-2026-40355, CVE-2026-40356 [2026 — NOT FOUND]Skills improved the structure of the analysis — explicit MITRE ATT&CK IDs (T1195.001), system-level krb5 C library CVEs surfaced, dependency-graph risk scoring, prominent training-cutoff warning. They did not change coverage. Every post-August-2025 CVE remained invisible.
“My training cutoff is August 2025. Today is June 27, 2026. Approximately 10 months of CVE disclosures are completely invisible to this analysis.” — Claude, with 817 skills loaded
Test 2 — NVD API query (the skill's Step 3)
The SBOM skill prescribes querying NVD 2.0 for each component. We executed it exactly. Results from live keyword search on 2026-06-27:
Query: "flask 2.3.0" → 0 CVEs
Query: "werkzeug 2.3.7" → 0 CVEs
Query: "requests 2.31.0" → 1 CVE (CVE-2023-32681) ← FALSE POSITIVE (fixed in 2.31.0)
Query: "krb5 0.7.1" → 0 CVEs
Query: "gssapi 1.8.2" → 0 CVEsCVE-2026-27205 was not found. Root cause: NVD keyword search doesn't match against CPE strings, and our SBOM had PURLs but no CPE identifiers. The skill correctly notes CPE queries are more precise — but methodology documentation doesn't guarantee correct execution when the inputs are incomplete.
Test 3 — pip-audit (the skill-recommended tool)
$ pip-audit -r /tmp/auditable-reqs.txt --no-deps
flask 2.3.0: ✓ CVE-2023-30861, ✓ CVE-2026-27205 ← 2026 FOUND
Werkzeug 2.3.7: ✓ 6 CVEs including CVE-2026-21860, CVE-2026-27199
requests 2.31.0:✓ CVE-2024-35195, CVE-2024-47081, CVE-2026-25645
krb5 0.7.1: ✗ FAILED — cannot build (no libkrb5-dev C headers)
gssapi 1.8.2: ✗ FAILED — samepip-audit found CVE-2026-27205 — because pip-audit makes a live OSV query at execution time. The skill pointed to the right tool, the tool ran, the live database had the 2026 data. But pip-audit could not audit krb5 or gssapi at all (missing C headers), and CVE-2026-40355 — the most dangerous finding, an unauthenticated remote crash — was missed by every skill-driven path.
Complete results: 4-way comparison
| CVE | Year | Sonnet / Opus | + 817 Skills (static) | NVD API (skill step) | pip-audit | Gadriel |
|---|---|---|---|---|---|---|
| CVE-2023-30861 | 2023 | ✓ | ✓ | ✗ | ✓ | ✓ |
| CVE-2024-34069 | 2024 | partial | ✓ | ✗ | ✓ | ✓ |
| CVE-2024-47081 | 2024 | ✗ | ✗ | ✗ | ✓ | ✓ |
| CVE-2026-27205 (Flask) | 2026 | ✗ | ✗ | ✗ | ✓ | ✓ |
| CVE-2026-21860 / 27199 | 2026 | ✗ | ✗ | ✗ | ✓ | ✓ |
| CVE-2026-25645 (requests) | 2026 | ✗ | ✗ | ✗ | ✓ | ✓ |
| CVE-2026-40355 (krb5 RCE) | 2026 | ✗ | ✗ | ✗ | FAILED | ✓ |
| CVE-2026-40356 (krb5 DoS) | 2026 | ✗ | ✗ | ✗ | FAILED | ✓ |
What this proves
1. Skills teach methodology, not CVE data
Loading 817 cybersecurity skills did not add a single post-cutoff CVE to what Claude could find. Skills are knowledge about how to perform security tasks — which tools to run, which APIs to query, which patterns to look for. They are not a substitute for the live databases those tools query. A medical textbook teaches a doctor how to diagnose pneumonia; it doesn't replace today's blood test.
2. Skills improved rigor, not coverage
With skills loaded, Claude's report was measurably better in structure — MITRE ATT&CK mappings, system-level CVEs surfaced, explicit risk scoring. The post-cutoff blind spot was identical. Skills added rigor inside the training window; they did not extend it.
3. The skill-recommended tools require execution — and still miss native libraries
When pip-audit was actually run, it closed the Flask / Werkzeug / requests gap. But executing a skill-recommended tool requires: installing it, running it, parsing output, acting on findings — every time, for every dependency change. And pip-audit still failed on krb5 and gssapi. CVE-2026-40355 — unauthenticated remote crash via NegoEx packet — was found by neither the NVD workflow nor pip-audit. Gadriel found it via OSV cross-ecosystem correlation against the SBOM's PURL data.
Gadriel vs. skills-armed Claude
| Step | Skills-armed Claude | Gadriel |
|---|---|---|
| Load skills / rules | 817 skills (manual install) | Built-in |
| Identify tools to run | Recommended pip-audit, grype | Built-in |
| Run SCA tool | Manual execution | Automatic |
| Handle native-library gap | Not solved | OSV cross-ecosystem |
| Generate SBOM | Not generated | SPDX 2.3 + CycloneDX 1.5 |
| Pre-commit gate | No | Yes (blocks on critical) |
| CVE-2026-27205 found | Only via manual pip-audit | Automatic |
| CVE-2026-40355 found | No | Yes |
The takeaway
The value of methodology skills is real, but bounded by execution. Skills correctly told Claude to run pip-audit. pip-audit found CVE-2026-27205. But every step from "skill suggests pip-audit" to "CVE blocked at the CI gate" requires a human in the loop — and still leaves native libraries uncovered.
Skills tell you to run the scanner. Gadriel is the scanner that runs itself.
Gadriel closes the loop automatically. It is not that Gadriel has better skills — it is that Gadriel is the tool the skills were pointing toward, executing deterministically, across ecosystems, every commit, with no manual install and no missing C headers.
Skills library: github.com/mukul975/Anthropic-Cybersecurity-Skills (817 skills, v1.3.0) · Gadriel Code Assurance v1.1.3 · OSV snapshot 270,336 advisories · 11 ecosystems · synced 2026-06-27.