A security audit skill based on Cloudflare's open-source security-audit-skill (MIT): evidence-first code review, sandbox-only execution, one Markdown report.
Updated November 2026.
Code security
Evidence-first
Sandbox only
Markdown report
Single agent
About This Skill
This skill makes your coding agent a defensive, evidence-first reviewer of a repository you own or may review: quick answers in chat by default, and a full audit with validation, an independent re-check and one REPORT.md only when you ask. Compared with the upstream, it condenses the multi-agent workflow into seven steps one agent can run, replaces the JSON schema and Node.js validators with Markdown, and adds an authorization check. Based on Cloudflare's open-source security-audit-skill (MIT). Not affiliated with or endorsed by Cloudflare.
1Download the skill (1 credit) and unzip the security-audit folder.
2Copy the folder into your skills directory: Antigravity ~/.gemini/config/skills/ (or .agents/skills/ in a project), Claude Code ~/.claude/skills/, Codex ~/.agents/skills/.
3Open a repository you own or are authorized to review.
4Ask a focused question for guidance mode, or explicitly request a full security audit to get a report under security-audit/<date>/.
Example Command
Run a full security audit of this repo and write a report
Core Capabilities
Guidance mode by default
Answers questions like whether an endpoint is safe in the chat, without creating any audit files.
Full audit on request
Recon, hunting by vulnerability class, validation, an independent re-check and a REPORT.md, only when you explicitly ask.
Six-slot findings
Actor, input, intended control, boundary crossed, affected party and result must all be filled before something counts as a finding.
Calibrated severity
Only confirmed findings get Critical to Info; unproven leads are marked needs-validation with the exact missing fact.
Strict safety rules
Runs target code only in an OS-enforced, network-less sandbox; no live probing, no installs, no copying secret values.
Smallest fix suggested
Each confirmed finding comes with a minimal fix and a regression test to add; the agent never edits your source during the audit.
Usage Examples
Input
Is the /api/invoices/export endpoint safe from cross-tenant reads?
Output
A guidance-mode answer: the traced path from route to query with file:line evidence, a verdict (confirmed, needs-validation or rejected), a severity only if confirmed, and the fix.
Input
Do a full security audit of this repo and give me a report.
Output
The agent confirms scope, report path and sandbox status, builds a coverage list, validates and re-checks candidates, then writes REPORT.md with findings, leads, hardening notes and coverage limits.
SKILL.md
---
name: security-audit
description: Defensive, evidence-first security review of a code repository the user owns or is authorized to review. Use it to answer security questions about the code or check one feature (guidance mode, the default). Run the full audit (reconnaissance, hunting by vulnerability class, validating each candidate, an independent re-check, and a report file) only when the user explicitly asks for a full security audit or a report.
---
# Security Audit
Find real vulnerabilities in a repository: places where a lower-trust actor can cross a trust boundary and cause a concrete result. Back every claim with source evidence, give it a calibrated severity, and propose the smallest fix. This is a defensive workflow. It reads code, and it runs code only inside a locked-down sandbox.
A single agent can run the whole workflow. If the agent can start parallel sub-agents, it may use them for hunting and re-checking. The rules stay the same either way.
## 0. Before anything else
1. **Confirm authorization.** Audit only code the user owns or is clearly allowed to review: their own repository, their employer's code that they work on, or an open-source project they are reviewing in good faith. If ownership is unclear, ask one question. If the user wants to attack a system they do not control, stop.
2. **Pick the mode.**
- **Guidance mode (default).** Use it for security questions ("is this endpoint safe?"), a check of one file or feature, triage of a report, or a methodology question. Use only the parts of this skill that apply. Do not create audit files. Answer in the conversation.
- **Full audit mode.** Use it only when the user explicitly asks for a full, comprehensive, or end-to-end audit, a pen-test of the codebase, or a written report. Run steps 1 to 7 below.
- If the request could mean either, ask: "Do you want a quick answer here, or a full audit with a report file?" Do not create any files before the user answers.
3. **Load the safety ru
Frequently Asked Questions
FAQ
How is this different from Cloudflare's security-audit-skill?
The wording is new throughout. The upstream six-phase multi-agent workflow is condensed into seven steps one agent can run, with sub-agents optional. The JSON schema, Node.js validators and budget gates are replaced by a Markdown finding format and one REPORT.md, and an authorization check plus a final checklist were added.
Is this the official Cloudflare security audit skill?
No. It is an Antigravity Skills edition based on Cloudflare's open-source project and is not affiliated with, sponsored by, or endorsed by Cloudflare, Inc. The official repository is https://github.com/cloudflare/security-audit-skill (MIT License).
Will it run my code or touch production?
It runs target code only inside an OS-enforced sandbox with no external network and strict limits. If that is not available, it stays read-only. It never probes live, staging or shared systems.
Where is the report saved?
By default under security-audit/<YYYY-MM-DD>/ in your project. The agent suggests keeping that folder out of public commits, and you can choose another path.
What does a download cost?
Each skill download costs 1 credit. New accounts start with 3 free credits.
Use these phrases to activate this skill in your AI coding assistant:
run a security audit on this repois this endpoint safecheck this feature for vulnerabilitiesfull security audit with a reportreview this code for auth bypasslook for cross-tenant data leakstriage this security report