Home / Add-ons / Security Audit Pack
Security Audit Pack — a source-first review workflow for coding agents
Short answer. 14 audit chapters, a report schema, two validators, a framework mapping, a report template, and a code review toolkit with a deterministic diff check and OSV dependency report. Built for source-code review, not a penetration test or live scan.
USD. If the file does not download, contact support for a replacement or a refund.
Re-download any time from this workspace entitlement.
See before you buy
Version 2026-11 · 53 files · 187.5 KB. SHA-256 2c5dca21e2261e1fac0c33e8257245a09f8c30de0c115c7b64ee1b48eab707cd.
The ZIP contains 14 chapter files plus a SKILL.md entry point, the report schema, two validators, the framework mapping, the report template, the MIT licence and the code review toolkit.
File list (53 files)
FRAMEWORK-MAPPING.mdLICENSES/security-audit-skill-MIT.txtREADME.mdREPORT-TEMPLATE.mdRUN-LOG.txtcode-review/CHECKLIST.mdcode-review/LICENCE.txtcode-review/README.mdcode-review/REPORT-TEMPLATE.mdcode-review/RUN-LOG.txtcode-review/THIRD_PARTY_NOTICES.mdcode-review/eslint.config.mjscode-review/patches/node-api-clean.patchcode-review/patches/node-api.patchcode-review/patches/python-api-clean.patchcode-review/patches/python-api.patchcode-review/prompts/build_prompt.pycode-review/prompts/llm-reviewer.mdcode-review/prompts/mock-response.jsoncode-review/prompts/request-example.jsoncode-review/prompts/schema.jsoncode-review/prompts/system.txtcode-review/qa/README.mdcode-review/qa/controls-last-run.mdcode-review/qa/last-run.mdcode-review/qa/run_controls.pycode-review/qa/run_qa.pycode-review/real-prs/manifest.jsoncode-review/real-prs/results-v2.jsoncode-review/review.pycode-review/reviews/REAL-PR-EVAL-V2.mdcode-review/sca.pyskill/AI-AND-LLM.mdskill/ATTACK-CLASSES.mdskill/CLIENT-SIDE.mdskill/CLOUD-AND-DEPLOYMENT.mdskill/DATA-ISOLATION-AND-LIFECYCLE.mdskill/DESKTOP-MOBILE-AND-LOCAL-IPC.mdskill/HUNTING.mdskill/MEMORY-SAFETY-AND-BINARY.mdskill/PROTOCOLS-RPC-AND-MESSAGING.mdskill/RECONNAISSANCE.mdskill/RESOURCE-EXHAUSTION-AND-AVAILABILITY.mdskill/SKILL.mdskill/SUPPLY-CHAIN-AND-RELEASE.mdskill/UPSTREAM-README.mdskill/VALIDATION-AND-REPORTING.mdskill/WEB-PROTOCOL-AND-AUTH.mdskill/report-schema.jsonskill/validate-coverage-ledger.cjsskill/validate-coverage-ledger.test.cjsskill/validate-findings.cjsskill/validate-findings.test.cjs
On this page · 15 sections
- Who it is for
- The problem it removes
- What you get
- How to use it
- What is in the pack
- The 14 chapters
- A sample finding
- What the validators reject
- Included: deterministic diff check and dependency report
- How it compares to a scanner
- What it is not
- How a review runs
- Why the framework mapping is labelled “indicative”
- What it costs and who it is for
- Licence
Who it is for
Engineering teams and consultancies reviewing repositories they own or are authorised to assess.
The problem it removes
- Security reviews depend on who ran them and what they remembered.
- Scanner output has no trust boundary and no evidence.
- A finding without a reproduction and a fix is a rumour.
What you get
- 14 chapters plus a SKILL.md entry point: auth, supply chain, cloud, AI/LLM, memory safety and more
- report-schema.json plus two dependency-free Node validators
- Framework mapping to MITRE ATT&CK, NIST 800-53 and SOC 2
- A report template with boundary, principal, evidence, reproduction and fix
- A code review toolkit: deterministic diff check, OSV dependency report, prompts and a review checklist
- MIT licence for the upstream skill, included in the ZIP
How to use it
1. Download and unzip
The ZIP contains the skill directory, the mapping, the template and the licence.
2. Run a guidance review
Ask your agent for a read-only trust-boundary review first; no file writes, no execution.
3. Validate the report
Run the two Node validators against findings.json; both are dependency-free.
curl -sS https://plugsky.com/api/addons/security/pack.zip \
-H "Authorization: Bearer $PLUGSKY_SESSION" \
-o plugsky-security-audit-pack.zip
What is in the pack
The pack is a working directory, not a slide deck. Your agent reads the chapters, applies the workflow to a repository, and writes findings against a schema the validators can check.
| Part | What it does | Why it matters |
|---|---|---|
| 14 chapters + SKILL.md | Reconnaissance, attack classes, web/auth, protocols, client-side, cloud, data isolation, supply chain, memory/binary, resource exhaustion, AI/LLM, desktop/mobile, hunting, reporting | Covers the boundaries a generic checklist misses |
| report-schema.json | Required fields for every finding | A finding without a boundary and evidence fails the schema |
| Two validators | Dependency-free Node scripts for findings and coverage | Catches empty claims before a report ships |
| Framework mapping | Indicative MITRE ATT&CK, NIST 800-53, SOC 2 references | Gives reviewers the language they ask for |
| Report template | Boundary, principal, evidence, reproduction, fix, priority | Turns an agent's notes into a reviewable document |
| Code review toolkit | review.py (diff rules), sca.py (OSV advisories), prompts and a checklist | A reproducible diff check and dependency report alongside the audit workflow |
The 14 chapters
| Group | Chapters | Typical source-review finding |
|---|---|---|
| Entry and identity | Reconnaissance, web protocol and auth, protocols/RPC/messaging | Auth bypass, IDOR, token confusion |
| Data and execution | Data isolation and lifecycle, memory safety and binary, resource exhaustion | Cross-tenant read, command injection, unbounded allocation |
| Platform and supply chain | Cloud and deployment, supply chain and release, client-side, desktop/mobile and local IPC | Public bucket, dependency confusion, insecure IPC |
| Applied | AI and LLM, hunting, validation and reporting, attack classes | Prompt injection, missing evidence, unvalidated severity |
A sample finding
{
"id": "SEC-001",
"title": "Tenant A can read Tenant B invoices by changing the id",
"severity": "high",
"boundary": "tenant",
"principal": "authenticated tenant A user",
"resource": "tenant B invoice records",
"evidence": ["app/Http/InvoiceController.php:84-96"],
"reproduction": "fixture tenants A and B; request B's invoice id with A's session",
"fix": "scope the query by workspace_id from the session, not the path id",
"frameworks": [{"name": "NIST 800-53", "ref": "AC-3"}],
"status": "confirmed"
}
That shape is deliberate: a reviewer can read the boundary, check the file, repeat the reproduction, and apply the fix without a meeting. A finding that cannot fill these fields is a needs-validation item, not a confirmed one.
What the validators reject
- Missing boundary or principal — a finding that does not name who crosses what.
- No source evidence — a scanner label is not evidence.
- A fix that only patches one instance — the schema asks for the class-level fix.
- Coverage claims without a ledger — the second validator checks that every reviewed area is recorded.
The 34/34 and 31/31 results are unit tests for the two validators. They show the checks behave as specified; they are not a measure of audit coverage or a guarantee that a repository is secure.
Included: deterministic diff check and dependency report
The pack now includes a small code review toolkit in code-review/. review.py parses a unified diff, maps added lines to new-file line numbers and runs 12 fixed rules. Tests, docs, fixtures, examples, vendor and generated files are classified as non-production and kept out of scored findings. Every scored finding carries a rule id, a file and line, and a remediation.
sca.py reads a changed package-lock.json, requirements.txt, go.mod or pom.xml and asks the free OSV.dev batch API for known advisories. It is a dependency report, not a reachability proof: a matching version is a prompt to investigate, not a confirmed exploit.
The toolkit also carries an LLM review prompt, a JSON schema, a human checklist and a report template. The defaults are deliberately conservative. On 20 merged pull requests from four permissive repositories (3,755 changed lines), the diff rules emitted 0 scored findings; 25 observations in tests and docs were excluded as non-production. The OSV positive control caught axios #11410 (follow-redirects@1.16.1, CVE-2026-116689). A zero-finding run is not evidence that a repository is secure.
This is not a hosted reviewer. There is no webhook, no auto-merge and no repository code execution. It is a local workflow that produces an artifact a reviewer can check.
How it compares to a scanner
A scanner reports patterns. This workflow reports a boundary violation with a principal, a resource and a path. That difference changes the output: a scanner can produce two hundred rows in a minute and none of them tell a reviewer whether the issue is reachable in this deployment. The pack forces the reachability question before a finding is confirmed.
It also changes the false-positive economics. A scanner's noise is cheap to generate and expensive to triage. The evidence requirement here is deliberately strict: file and line, the configuration that makes the path live, and a safe reproduction with fixtures. If any of those is missing, the item is marked needs validation rather than counted as a finding. A report with ten confirmed findings and four open questions is more useful than one with sixty rows and no owner.
The trade-off is time. A guided review of a service takes hours, not seconds, and it depends on the reviewer's judgement. The pack is not a replacement for a scanner; it is the layer that decides which scanner output matters and what to do next.
What it is not
It is a defensive, source-first code review. It does not scan a live service, test availability, or run a penetration test. The workflow explicitly requires a sandbox before any target-controlled code executes, and reports a missing sandbox as a blocker instead of improvising one.
That boundary is the product. A pack that promises a pentest creates a liability; a pack that produces evidence-backed source findings creates a review.
How a review runs
- Scope. Name the repository, the commit, and the trust boundaries that matter: tenant, user, process, network, model context.
- Guidance first. Ask for a read-only review of one module. No writes, no execution, no report directory.
- Evidence. Every candidate finding must name the principal, the resource, the file and line, and the path that makes it reachable.
- Validate. Run
node validate-findings.cjs findings.json. A finding that fails the schema is not ready. - Report. Use the template. Separate confirmed findings from needs-validation items.
The pack ships with the upstream skill's own tests. We ran both validators before shipping: 34/34 and 31/31 pass. The run log is in the ZIP.
Why the framework mapping is labelled “indicative”
MITRE ATT&CK is a behavioural model. NIST 800-53 and SOC 2 are control models. There is no one-to-one crosswalk, and a finding that maps to a control does not make a system compliant. The mapping exists so a reviewer can find the right conversation, not so a report can claim an audit opinion.
What it costs and who it is for
It is $49 once. It earns its price when a team reviews repositories repeatedly and wants the same structure each time: an agency doing pre-delivery reviews with the client's authorisation, a platform team reviewing a service before launch, or a maintainer preparing a security questionnaire response. It is not a substitute for a specialist when the system is high-risk, and it is not legal advice.
Licence
The audit skill is MIT, copyright Cloudflare, Inc. The licence text is in LICENSES/. Plugsky is not affiliated with Cloudflare and does not imply endorsement. Our framework mapping and report template are original and may be used inside your organisation on repositories you own or are authorised to review; they may not be resold as a standalone product.
Questions
What exactly do I get?
A ZIP with 14 audit chapters, report-schema.json, two validators, a framework mapping, a report template, the MIT licence, and a code review toolkit: deterministic diff check, OSV dependency report, prompts and checklist.
Is this a penetration test?
No. It is a defensive, source-first code review. It does not scan live systems, probe endpoints or test availability.
Which agent does it work with?
Any agent that reads a directory of skills: Claude Code, Codex, Cursor, Gemini CLI, OpenCode. The validators need Node.
Are the validators tested?
Yes. The validators pass 34/34 and 31/31 unit tests; the run log is in the ZIP. Those results cover the validators, not the depth of an audit.
Can I use it commercially?
Yes, inside your organisation on repositories you own or are authorised to review. Resale of the pack itself is not permitted. Agencies should contact us for client-repository terms.
Sources
- Cloudflare security-audit skill (MIT) — checked 22 November 2026
- MITRE ATT&CK — checked 22 November 2026
- NIST SP 800-53 — checked 22 November 2026
Related add-ons
Prompt Cost Linter
Static analysis for seven LLM cost anti-patterns: full context resent every turn, missing prompt caching, flagship models on simple tasks, unbounded output, retry storms, batchable loops and needless re-embedding.
Prompt Pack
Forty production prompts across extraction, classification, support and coding, each with a fixed output shape, two or more eval cases and a runner that checks them against the API.
Stripe checkout. The download appears on this page and in Dashboard → Billing → Add-ons as soon as payment completes.
$49 — buy onceCompare add-ons