Coinjar attack lab and audit report
Coinjar is a small practice backend that sells, grants and pays out coins. Its first version loses money in 17 ways. When this page was built, each attack was sent to that version and to a fixed one. The requests and answers below are the recorded ones.
- Attacks
- 17, each run on both versions
- Audit findings
- 24, 20 proved with live requests
- Planted bugs found
- 17 of 17
- Audit run
- 2026-11-13, Claude Code
- Page built
- 2026-11-16, Node v22.23.4
- Skills version
- 1.3.5, commit 78670a3
Loading the recorded run
Attack lab
Pick an attack to see the requests it sent, what the first version and the fixed version answered, and the code behind each answer.
All 17 results in one table
Each attack ran against a fresh copy of both apps on 127.0.0.1. The apps are in playground/, and npm test runs the same 17 attacks on your machine.
Audit report
An AI coding agent ran the money-audit skill on a copy of the first version, without the answer key, the fixed version or the tests. Each of the 17 planted bugs matched a different finding, and 23 of the 24 findings point at a planted bug. Every finding names a file and line, quotes the code, and has a request proof or a code trace.
- Findings
- 24
- Rated critical
- 11
- Proved with live requests
- 20
- Planted bugs found
- 17 of 17
Coverage: what was checked in each of the 10 classes
Notes the agent added below the findings
How the app was run during the audit
One run with Claude Code. The files are in the repository as they were written: findings.json, REPORT.md and the saved probe outputs. Coinjar was written for this repository, so the result says how the skill does on a known target, not on your code.
Skills
One skill finds the bugs. The other 10 build the parts that prevent them, in your stack and with your conventions. Each skill is a folder of plain Markdown that your agent reads when a task calls for it.
Install
The skills follow the open Agent Skills format, so they work with Claude Code, Cursor, Codex, GitHub Copilot, Gemini CLI, OpenCode, Goose, Amp and other agents that read it.
npx skills add moonetix/revenue-skills
npx skills add moonetix/revenue-skills --skill money-audit
/plugin marketplace add moonetix/revenue-skills /plugin install revenue-skills@revenue-skills
codex plugin marketplace add moonetix/revenue-skills
Then ask in plain words
- Audit this project for ways users could get free credits.
- Our payment webhook sometimes credits an order twice. Find out why and fix it.
- Add a referral program that cannot be farmed with fake sign-ups.
Or measure your agent on Coinjar first
git clone https://github.com/moonetix/revenue-skills cp -r revenue-skills/playground/vulnerable coinjar # ask your agent: "Run a full money audit of ./coinjar with the money-audit skill." node revenue-skills/playground/score.mjs coinjar/money-audit/findings.json
The probe that produces the proofs only talks to 127.0.0.1 unless you name another host, and it needs Node 20 or newer. The skills installer needs Node 22.20 or newer. Nothing is installed besides Markdown and a few small scripts.
Maintained by Moonetix, an affiliate network.