Sample audits · Workstation audits, October 2026

cloudflare/security-audit-skill

A published security-audit procedure for AI agents: guidance pages, report schema and two validators.

Summary card for this audit: 10 findings after review, 0 High, 0 Medium, 8 Low, 2 Info
Auditedcloudflare/security-audit-skill at commit c1c8a8c1471069fb0e188eeaff69b8e8db6564a8
Date22 October 2026
How it ranlocal run on our workstation, full audit, Standard review
Verdict after reviewPass with notes (Fail if a High finding remains after review, otherwise Pass with notes)

Related: Node.js API documentation, part 1 of 2 (October 2026) · Node.js API documentation, part 2 of 2 (October 2026)

0
High after review
0
Medium after review
8
Low after review
2
Info after review
0
excluded on review
How to read this page

Each finding keeps the number it has in the audit report. The rating shown first is the one after review; the first automated rating is listed with it. 0 findings were first rated Medium or High; 0 of them are Medium or High after review. Text marked "From the report" is quoted from the audit report; fixes are suggestions and were not tested. Code locations link to the file and line at the audited commit.

Low after review (8)

1. Low Report text copies target-derived strings with no escaping step
First rating: Low · Reviewed rating: Low · Review: confirmed
From the report
Evidence

Evidence: For each confirmed medium, high, or critical record, copy the complete source path and target-neutral local reproduction:

Why it matters

The reports are built from text taken from the audited repository (paths, snippets, reproductions). Nothing says to fence or escape it, so markup or links planted in the target can render when the report is opened in a viewer or posted to a tracker. Rated LOW because no rendering path is shown in the pack.

Suggested fix, not tested

State that target-derived text goes into code fences or is escaped, and that raw HTML and link targets from the target are neutralised.

2. Low The findings validator has no per-object field cap or total work budget
First rating: Low · Reviewed rating: Low · Review: confirmed
From the report
Evidence

Evidence: const LIMITS = Object.freeze({ inputBytes: 5 * 1024 * 1024, nestingDepth: 64, arrayItems: 1000, ... validationErrors: 100 });

Why it matters

It caps bytes, depth, array length and errors, but not fields per object or a total value count, unlike the ledger validator. The input is the parent run's own output and the 5 MiB cap bounds the cost, so this is a consistency gap.

Suggested fix, not tested

Add objectFields and total-value limits as validate-coverage-ledger.cjs has.

4. Low Nothing checks the two validated files against each other or against the report
First rating: Low · Reviewed rating: Low · Review: confirmed
From the report
Evidence

Evidence: Validator success proves format and ledger consistency only.

Why it matters

Each validator checks its own file. The page admits this, but no step checks that every ledger result fingerprint appears in findings.json or that the report matches the records; only the parent's discipline enforces it.

Suggested fix, not tested

Add a small cross-check, or a documented parent-side step, covering both directions of the fingerprint match.

5. Low Security-critical input helpers are duplicated across the two validators
First rating: Low · Reviewed rating: Low · Review: confirmed
From the report
Evidence

Evidence: function readFileWithinLimit(file) { const noFollow = fs.constants.O_NOFOLLOW; ... (same body at validate-findings.cjs:604)

Why it matters

The file reader, error list, quoting helper and path check exist twice. Only the regexes and a path corpus have parity tests, so the copies can drift apart.

Suggested fix, not tested

Move readFileWithinLimit, safeQuote, the error list and the path helpers into one shared .cjs module.

6. Low Both validators refuse every input on Windows, and the docs do not say they are POSIX-only
First rating: Low+ (potentially higher) · Reviewed rating: Low · Review: confirmed
From the report
Evidence

Evidence: if (!Number.isInteger(noFollow) || noFollow === 0 || !Number.isInteger(nonBlock) || nonBlock === 0) { throw new SafeInputError("OS no-follow and nonblocking input protection is unavailable")

Why it matters

Node on Windows defines neither O_NOFOLLOW nor O_NONBLOCK, so both validators reject all input there (validate-coverage-ledger.cjs:721 has the same code). The workflow says a run ends only when both validators pass, and the README requirements list only Node.js. The run can still be closed as incomplete and the failure is loud, so nothing is lost.

Could be higher if: Windows is a supported platform for people running the skill.

Suggested fix, not tested

Document the validators as POSIX-only, or add a Windows path that opens read-only and compares lstat and fstat identity before reading within the size cap.

7. Low The procedure never says that audited-repository content is untrusted data
First rating: Low+ (potentially higher) · Reviewed rating: Low · Review: confirmed
From the report
Evidence

Evidence: 2. architecture.md verbatim. (the required hunter prompt parts list has no untrusted-content statement)

Why it matters

A search for 'untrusted', 'prompt injection', 'not instructions' and 'as data' in the main procedure, hunting, reconnaissance and reporting pages found no statement for target content. Reconnaissance output is pasted verbatim into later hunter prompts, so text planted in the target could steer agents or suppress findings. Rated LOW as an omission with no demonstrated exploit.

Could be higher if: Hunters run unsandboxed or with write or network tools on hostile repositories.

Suggested fix, not tested

Add a trust-boundary paragraph to the main page and to every hunter, verifier, critic and reconnaissance prompt: target content is data, embedded directives are reported as findings, and architecture notes derived from it are untrusted too.

8. Low No rule to truncate or mask secrets found in the target when writing evidence and reports
First rating: Low+ (potentially higher) · Reviewed rating: Low · Review: confirmed
From the report
Evidence

Evidence: Are there any hardcoded passwords, API keys, tokens, or secrets in the source? (grep for password, secret, apikey, token, Bearer, -----BEGIN, ...)

Why it matters

The procedure sends hunters to find credentials, including ones removed from git history, but no page tells them to mask the value in evidence, payloads or the report. A full secret could then be written into files that are shared or committed.

Could be higher if: Reports are shared or committed with the target repository.

Suggested fix, not tested

Add a rule that findings and report text give the file, line and type of a secret and at most its first 4 characters.

9. Low Safe-path check rejects valid repository paths, so real findings can fail validation
First rating: Low · Reviewed rating: Low · Review: confirmed
From the report
Evidence

Evidence: !/[ .]$/u.test(segment) && !WINDOWS_RESERVED_COMPONENT.test(segment) (and value.includes(":") at line 474)

Why it matters

Legal POSIX names such as aux.c, nul.go, a file with ':' in its name, or one ending in '.' cannot be cited in a trace, evidence or code change. The validator fails loudly, so the finding must be cited another way. validate-coverage-ledger.cjs isSafeRelativePath (line 377) has the same rule.

Suggested fix, not tested

Validate only traversal and absolute forms, or offer a documented escape for such names.

Info after review (2)

3. Info No schema or version field in the run files although prior-run logic depends on compatibility
First rating: Low · Reviewed rating: Info · Review: rated too high
From the report
Evidence

Evidence: Record missing or incompatible ledgers instead of treating them as empty coverage.

Why it matters

'Incompatible' is never defined, and the schema uses additionalProperties:false, so a future field change silently invalidates earlier runs' findings.json.

Suggested fix, not tested

Add a required schema_version to the ledger, the findings contract and run-metadata.json, and define what incompatible means.

10. Info A UTF-8 byte order mark makes valid input fail with an unhelpful message
First rating: Low · Reviewed rating: Info · Review: rated too high
From the report
Evidence

Evidence: const UTF8_DECODER = new TextDecoder("utf-8", { fatal: true, ignoreBOM: true });

Why it matters

With ignoreBOM true the BOM stays in the string and JSON.parse fails with 'invalid JSON syntax'. The ledger validator (line 88) does the same. The failure is loud and the fix is to re-save the file, so nothing is lost.

Suggested fix, not tested

Set ignoreBOM to false, or strip U+FEFF before parsing.

Want this for your code?

Upload a ZIP or link a public GitHub repo, pick the areas and the depth, and get findings by severity with fixes.