Attack your own AI-built app first: Cloudflare's audit skill and Strix

My take: If an AI wrote your app, let another AI attack it before a stranger does. Only on code you own, and with an eye on the token bill.
What to know
- Cloudflare's security-audit skill (MIT license, 73,656 GitHub stars on November 16, 2026) turns a coding agent such as Claude Code into a code auditor. In six phases, it maps the app, hunts flaws by attack type, hands every candidate to a fresh agent whose only job is to disprove it, then writes a report. It needs an agent that can run sub-agents, Node.js, and a sandbox cut off from the network before it runs your code; without one, findings stay unconfirmed.
- The 7,245 findings Cloudflare published on June 18, 2026 came from a larger internal system grown out of this skill, which Cloudflare had not published on GitHub as of November 16, 2026.
- Strix (Apache 2.0, 75,926 stars on November 16, 2026) attacks the running app instead: from a Docker sandbox, teams of agents run your code or hit your site, try to exploit what they find, and return a working proof for each flaw. The command-line tool is free with Docker and your own model API key; Strix Cloud and Enterprise are paid. Among recommended models, its README marks Claude Opus 5.5 as needing Anthropic's Cyber Verification Program.
Spotted on @henriexploria and @0xloucash, in French.
What changes
Install the skill with npx skills add https://github.com/cloudflare/security-audit-skill --skill security-audit, then ask your agent for a security audit of the repository. Cloudflare says one run found about half of what repeated runs found: run it more than once.
Point Strix only at code and sites you own. Its README says testing a system without written permission is illegal in most jurisdictions. Both run several agents at once, each spending model tokens: on a subscription, that eats your limits, so start with one small repository. Then read each finding before you fix it: a report is a lead, not a verdict.
Source: Grant Bourzikas (Cloudflare), “Build your own vulnerability harness”, June 18, 2026
Drafted with Claude, Anthropic's AI, from my take. Reviewed and owned by me.