Fails closed
A crash, an invalid boundary, a replay that errors or a server older than your code blocks the stop. None of them counts as a pass.
notyours: COULD NOT CHECK … this is not a passnotyours signs in to your app as several users and plants a canary token in each account. If one user's token reaches another user's browser, your coding agent is blocked from finishing and gets the evidence; after 3 failed attempts it escalates to you.
Passport no. NY-ALICE-JOTTERSEED2A
Passport no. NY-ALICE-JOTTERSEED2A
Jotter is the bundled demo app. Its bugs are seeded on purpose.
Recorded on the demo app: Claude Code adds a feature, the Stop hook replays the locked boundaries, the agent fixes the server, and the replay passes.
From docs/loop-run-1. A passing Stop prints nothing; the HELD lines summarise its exit 0.
A crash, an invalid boundary, a replay that errors or a server older than your code blocks the stop. None of them counts as a pass.
notyours: COULD NOT CHECK … this is not a passThe guard hook blocks the agent's edits to locked boundaries, Kane recordings, run results and config. A guardrail for a cooperating agent, not a sandbox.
notyours: tests and boundaries are read-only for the agent.Once a person approves the start command on this machine, the Stop hook restarts your app first, so a replay never judges old code.
app: { start: "node server.js", port: 3000 }sed -i, scripts and codegen don't fire edit hooks. Each Stop pass also compares the boundary's files with the last snapshot.
Access-control bugs are small: a query without owner_id = ?, a list that returns every row. Unit tests run one user at a time, so they rarely see them.
of the applications OWASP tested had some form of broken access control.
OWASP Top 10:2025, A01, backgroundA researcher scanned 1,645 apps built on one AI app builder and found 170 whose database rules exposed user data.
CVE-2025-48757, the researcher's statementReal browsers, real sign-ins, and a verdict rule that gives the same answer on every replay.
notyours map scans Express and Next.js routes and writes draft boundaries. You, or /notyours-map in Claude Code, finish them.
Kane signs in as Alice and Bob in separate Chromes. Alice's record carries a canary token; Bob visits it. The token decides.
notyours lock turns each proven boundary into a recorded Kane test. Later checks replay it, with no authoring runs.
The Stop hook replays changed boundaries each time the agent tries to finish. A leak blocks the stop and hands the agent the evidence.
Every probe also checks that the owner and the viewer can see their own tokens. A probe that couldn't have found anything doesn't count.
The owner's token appeared in the viewer's rendered page.
The token reached the viewer's browser in a network response, even though the page didn't show it. Anyone can open DevTools.
The viewer never received the owner's token, and both positive controls passed.
A positive control failed, so the probe proves nothing. It is never counted as a pass.
Kane's pass/fail never decides; the token does.
An excerpt of the sample report from the demo app, with emails redacted. The same findings go to your coding agent.
2 boundaries leaked out of 4. Bob saw Alice's data (1); Bob got a Pro feature for free (1).
| LEAK visible | b1-notes-cross-user | Bob tried to open data Alice created. | NY-ALICE-JOTTERSEED2A on screen at /notes/1001 |
| LEAK in payload | b3-export-plan-gate | Bob (free plan) tried a Pro-only action. | NY-BOB-JOTTERSEED2B in the /api/export response |
| HELD | b2-notes-guest | A signed-out guest tried to open data Alice created. | Guest never received Alice's token |
| HELD | b4-bob-sweep | Bob swept every page, looking for anything Alice created. | Bob never received Alice's token |
A fresh Claude Code session built a notes app with share links and a Pro export, from a neutral prompt with no security instructions. notyours tested a copy of it in three experiments and found 0 leaks.
Cross-user, guest, plan-gate and sweep, on the app as Sonnet 5.5 built it.
"Note pages feel slow. Add a simple in-memory cache…" given to Sonnet 5.5.
The same cache prompt, given to Haiku 5.5.
The experiments also tested notyours itself. Fixed: file downloads, a strict CSP blocking the oracle's fetch(), the Stop hook judging a stale server, and failed login replays for personas still signed in.
What didn't happen: neither model took the bait, so there was no real regression to catch. The full catch-and-fix loop has only been shown on the demo's seeded bugs.
The experiments and the fix-day re-runs cost about 559 Kane credits over 35 authoring runs. Read the summary.
notyours checks the boundaries you declare, in the running app. It isn't a penetration test.
--allow-host and confirm.The quick demo finds one real leak in the bundled app. Measured: 4 min 48 s and 61.69 Kane credits. It prints its estimate before it starts.
npx notyours doctorChecks Node, Chrome, kane-cli and its login, credits and port 3077.npx notyours demo --quickOne real leak in the bundled demo app, about 5 minutes.npx notyours initIn your app's folder: finds the app, asks you to approve its start command, writes boundary drafts.npx github:Jasekeee/notyours doctorThe same commands, built straight from GitHub on the first run. Use it only if npm is unavailable.npx github:Jasekeee/notyours demo --quickOne real leak in the bundled demo app, about 5 minutes.npx github:Jasekeee/notyours initIn your app's folder: finds the app, asks you to approve its start command, writes boundary drafts.# macOS brew install lambdatest/kane/kane-cli # anywhere else npm i -g @testmuai/kane-cli kane-cli login
No. Kane drives the browsers, and the verdict comes from the canary token. Nothing is judged by an AI.
Kane credits from your TestMu AI account. Measured: the quick demo used 61.69 credits (4 min 48 s), and a replay of four locked boundaries used 3.4 credits (170 s). run --dry-run prints an estimate without starting Kane, and the CLI caps authoring runs per project per day (default 20).
Yes, through Kane. Your app runs on localhost, but every Kane run belongs to your TestMu AI account and gets a Test Manager link, and Kane's agent decides its steps from what the personas' browsers show. Treat what your test personas see, and the test accounts' credentials, as shared with TestMu AI: use test accounts and test data. notyours itself sends nothing; the oracle runs on your machine. Evidence packs hold your account email and share links; init keeps them out of git, so don't publish them.
The route scan reads Express and Next.js code. Any app that runs on a URL can still be tested; you write its boundaries by hand.
The guard hook blocks the agent's edits to locked boundaries, recordings, run results, budget and config, and the Stop hook fails closed. It's a guardrail for a cooperating agent, not a sandbox: an agent with shell access could find spellings the guard doesn't anticipate.
Run the demo in about five minutes, then point it at your own app.