DappSentinel

Your dApp doesn't ship until Kane says it's safe.

DappSentinel is a pre-launch safety gate for dApp websites. Kane CLI checks the site in a real browser for the crypto mistakes you can't see on screen. Claude Code fixes what Kane catches, and Kane checks again until every check passes.

  • No RPC API keys in traffic
  • No testnet endpoints
  • Right contract addresses
  • Right chain ID
  • Scripts from trusted hosts
  • No keys in browser storage
  • Swap flow works
Gate closed: 6 of 7 checks failed

The contracts were fine. The websites weren't.

Some of Web3's costliest attacks never touched a smart contract. They came through the website users trusted.

Dec 2021

BadgerDAO

A malicious script was injected into the website through a compromised Cloudflare API key.

Source: The Block
~$120M
Feb 2025

Bybit, via Safe{Wallet}

Tampered JavaScript in the Safe{Wallet} web interface altered a multisig transaction.

Source: The Block · SlowMist
~$1.5B

About $1.62 billion, through dApp websites. Figures are approximate, at the time of each attack.

What Kane checks

Kane drives a real Chrome and uses the site like a person. DappSentinel reads the evidence Kane records: every request, every script, everything left in storage.

Leaked RPC keys

Provider API keys (Alchemy, Infura, QuickNode…) in the requests the page sends.

Kane's network recording

Testnet in a mainnet build

Sepolia, staging or local endpoints left in a production site.

Kane's network recording

Wrong contract address

A token address on the page that's off by even one character.

Kane reads the page

Wrong chain

A network badge that says Sepolia when you meant mainnet.

Kane reads the page

Untrusted scripts

Code loaded from any host you haven't allowed. The BadgerDAO problem.

Kane's network recording

Keys in browser storage

Private keys or seed phrases saved where any script on the page can read them.

The browser Kane drives

Plus a seventh check that the main flow still works: Kane runs the swap from start to finish, so a fix can't pass by breaking the page. That run is also the one whose traffic the network checks read. On a live dApp DappSentinel doesn't know, it runs the five checks that need no inside knowledge: the four above that read traffic and storage, and a check that the page loads.

Kane catches it. Claude fixes it. Kane checks again.

Kane CLI

Opens the site in a real browser, runs the main flow and records the network traffic.

DappSentinel

Scans what Kane recorded and watches browser storage live. Writes the report.

Claude Code

Fixes the site. A Stop hook won't let it finish while any check fails.

↺ After every fix, the hook re-runs Kane. The loop ends only when every check passes, and the gate opens.

The loop closes on GitHub too. In PR #5, nobody touched anything: Kane failed the pull request 6 of 7, Claude Code fixed the site on GitHub and pushed the fix, and the pull request's own gate passed 7 of 7. Only demo-dapp/ may change, and it stops after 3 rounds.

Why Kane, and not a scripted browser

A script can only use the site it was written for. Kane takes a plain-English goal and works the page out for itself, so the same checks run on any dApp with no per-site code.

Aave, never seen before

Kane dismissed a notice banner on its own, then confirmed the market is USDC on Ethereum and shows a supply APY and the total supplied.

Lido, never seen before

Kane filled in the stake form and confirmed it quotes 1 ETH = 1 stETH, shows an APR and runs on Ethereum. 6 of 6 passed.

Three jobs

Kane decides what passes, produces the evidence the network and storage checks read, and tells Claude Code what's wrong.

Verdicts, evidence, feedback

The agent can't cheat

The only way to green is fixing the site. Changing the referee doesn't work.

Claude Code's attempt to edit a check, blocked with the message DappSentinel: blocked
Edit a check: blocked. A Claude Code hook stops edits to the checks, the config and the gate.
Terminal showing: Refused, the referee was changed
Change it any other way: refused. The gate compares the checks with the last commit, and with the target branch in a pull request. On GitHub: PR #6 edited a check and the gate refused to run.

Every run is public

Real GitHub Actions runs, unfiltered. Same input, same verdict.

InputResultRuns
Demo dApp with six planted mistakes6 of 7 failed, every time1, 2, 3
The fix, in PR #17 of 7 passed, every time1, 2
Broken, then fixed by Claude Code on GitHub, in PR #3 and PR #56 of 7, then 7 of 7 after Claude's fix, both times#3 red, #3 green, #5 red, #5 green
A broken change, in PR #26 of 7 failed, merge blocked1
Mutation test (known-answer): the fixed site plus 7 copies, each with one new mistake the checks were never tuned for (full record)8 of 8 judged correctly, both times1, 2
An edited check, in PR #6Refused to run, auto-fix not started1
Lido, with Kane using the stake screen6 of 6 passed1
Aave, with Kane using the USDC market5 of 6: Kane's screen check passed; same script hosts as before1
Lido (live, mainnet)5 of 5 passed1, 2, 3
Aave (live, mainnet)4 of 5: scripts from app.family.co (Aave's own wallet infrastructure) and Cloudflare Insights1
CoW Swap (live, mainnet)3 of 5: an Alchemy RPC URL carrying a key, and scripts from 11 ad and analytics hosts1

Mutation test results: the clean site passes 7 of 7, and each of 7 copies with one new mistake fails exactly its own check, on both runs

A flagged key or script host is a prompt to look, not proof of an attack: many dApps ship keys that only work from their own domain. Keys are masked in every report. The demo has 7 checks; live dApps get the 5 that need no inside knowledge of the site. PR #2's run is also the third run of the broken demo. The full list, including the runs that don't count and why, is in the README.

Install in one line

In GitHub Actions

Add your TestMu AI username and access key as repository secrets, then one step. The pull request fails while any check fails.

- uses: Guzman0/DappSentinel@v1.1.0
  with:
    url: https://staging.yourdapp.xyz
    lt-username: ${{ secrets.LT_USERNAME }}
    lt-access-key: ${{ secrets.LT_ACCESS_KEY }}

On your machine, with Claude Code

Run the gate yourself, or run /sentinel in Claude Code and let it fix what Kane finds.

# Node 18+, Chrome, a TestMu AI account
npm install -g @testmuai/kane-cli
kane-cli login
git clone https://github.com/Guzman0/DappSentinel
cd DappSentinel && npm install
npm run sentinel

Try it on the live demo dApp: a deliberately vulnerable swap page with all six mistakes left in, built as the scan target.

What it doesn't do

  • It isn't a smart contract audit. It checks the website, not the contracts behind it.
  • It can't catch every frontend compromise. It sees what the browser does while Kane uses the page. Code that only wakes up for certain wallets, like the code in the Bybit attack, may stay quiet during a check.
  • It keeps only a redacted report. Keys are masked everywhere in it, and Kane's raw network recording is never uploaded unless you opt in. Security and data handling.
  • Some sites block automated browsers. When Kane can't reach the real page, DappSentinel reports the checks as inconclusive, never as clean.