Check the code before you trust it
Coming soon to Android and iOS.
A QR code is a URL you cannot read. QR Protector resolves it the way your phone would, inspects where it actually lands, and gives you one plain-English verdict — before you ever open the link.
The problem
Every other link on your phone shows you where it goes. A QR code shows you a square. By the time it has resolved you are already on the page, and if that page is a convincing login form the decision has been made for you.
Worse, the destination can change after the sticker is printed, and it can serve you something different from what it serves a security scanner. A short link that looks clean from a data centre can land somewhere else entirely when it thinks it is talking to a phone.
The attack is not the code. The attack is everything that happens between the code and the page.
How it works
Every scan runs the same pipeline. The expensive AI stage only ever corroborates what the deterministic checks already found — it is never allowed to invent a risk on its own.
Camera or gallery. The scanner guards against the same code being processed twice while a scan is already in flight — an unglamorous fix that is documented as the first critical one in the architecture notes.
scanner_screen.dartUp to ten hops, following each one manually rather than letting the HTTP client do it, and recording the source, destination and status of every hop. The app renders the whole chain rather than just the endpoint.
redirectResolver.jsThe same URL is resolved in parallel under desktop, iPhone and iPad user-agents. If the registrable domains disagree, the destination is cloaking — showing one thing to scanners and another to phones. A device that dead-ends is recorded as exactly that, never quietly rounded up to 'consistent'.
domainDiverged / deadEndDeviceGoogle Web Risk runs against the final URL and each intermediate hop in parallel. A lookup that fails is recorded as unverified — it is never coerced into a pass.
Issuer, common name, hostname and SAN match, and chain trust for the final hop — the one you would actually have landed on.
certClient.jsPassword fields, payment fields, meta-refresh targets, off-site link domains, and Android package download intent. Most importantly: the registrable domain that each form actually submits to, which is the thing a screenshot can never tell you.
contentIntent.jsA login form on one domain posting credentials to an unrelated one is a strong mismatch. A certificate for a different registrable domain is a weak one. Payment processors and identity providers are allow-listed, because legitimately posting off-site is exactly what they do.
brandCorrelation.jsPunycode and mixed-script homographs, raw-IP hosts, shorteners, and around twenty reputation-laundering hosts where anyone can get a subdomain on a trusted parent. Domain age is fetched over RDAP and cached — and only when a form is present, because that is the only time it changes the answer.
domainIntel.jsFor premium accounts: real screenshots at three viewports, text and logo detection, then a Gemini pass with the screenshots attached. Its system prompt casts it explicitly as a corroborating voice — deterministic checks run alongside it, and it is told not to invent risk.
"a CORROBORATING voice"Dangerous requires more than one thing agreeing: an app download from a non-store domain, or a brand mismatch together with a credential form, or device divergence together with a form. A clean shortener that lands on a valid, reputable domain is not flagged — because crying wolf is how a security app teaches people to ignore it.
deriveBackendStatus()The explainer returns the final status, the named branch that decided it, every input and every gate boolean. It was built to diagnose false positives rather than merely to ship verdicts.
RiskGatingPolicy.explain()Why it works this way
10
Redirect hops
each one resolved, badged and shown to you
3
Devices at once
desktop, iPhone and iPad — because a link can serve each of them something different
~20
Laundering hosts
trusted parent domains anyone can get a free subdomain on
Honest by design
Plenty of security tools flag everything faintly unusual and call it vigilance. That just teaches you to ignore them. QR Protector only says dangerous when more than one independent check agrees — and when it isn't sure, it tells you that instead of picking a side.
The reputation lookup didn't come back. That isn't the same as clean, so we're not going to tell you it is.
Unverified — not a pass
Everything inside
The whole journey, not just the destination — with per-hop reputation and certificate badges, so you can see exactly where a link stopped being trustworthy.
Every hop shown with its own reputation and certificate — never just the endpoint.
Every verdict can show the branch that decided it, every input it used, and every gate boolean. Built to diagnose its own false positives.
Shows the branch that decided it — never just a colour.
Everything inside
Resolves the same URL as a desktop, an iPhone and an iPad at the same time. Destinations that differ by device are the clearest signal there is that something is hiding.
Compares registrable domains rather than full URLs, so ordinary per-device paths don't produce false alarms.
A device that dead-ends is recorded as a dead end, not as agreement.
A device that dead-ends is recorded as a dead end — never as agreement.
Reads the destination's HTML for password fields, payment fields and app-download intent — and for the domain each form actually posts to.
Where a form submits is the single most useful signal on a phishing page, and it is invisible in a screenshot.
Reads where the form actually posts — never just how the page looks.
Punycode and mixed-script look-alikes, raw-IP hosts, shorteners, and the trusted parent domains that anyone can get a free subdomain on.
Issuer, common name, hostname and SAN match, and chain trust — checked on the hop you would actually have landed on.
Local, searchable and filterable, stored on the device rather than in an account.
History stays on your device — never in an account.
A ninety-question quiz on QR scams with scoring and attempt history, plus collapsible reference cards on how the attacks actually work.
A standalone command-line companion that scans Markdown and text for leaked secrets, prompt injection, dangerous commands and disguised links.
Human, JSON and SARIF output, so it drops into a CI pipeline.
An optional visual mode renders the document and looks for hidden or deceptive text.
Live scanning, or check a code someone has sent you as an image before you act on it.
See it in action
Architecture
A Flutter client that holds no secrets, a Node backend on Cloud Run that holds all of them and does the analysis, and server-side entitlements so the expensive pipeline cannot be unlocked from the client.
Background
QR Protector exists because a QR code is the only link on your phone you are asked to trust completely blind. The whole product is an attempt to give you back the two seconds of judgement that every other link affords you.
The architecture follows from one decision: the app holds no secrets. Every API key lives behind a Node backend on Cloud Run, and the client authenticates with a Firebase ID token and asks the backend to do the work. That also made the security posture testable, because there is exactly one place where the expensive, abusable operations happen.
The analysis itself is deliberately boring before it is clever. Seven deterministic checks run first — redirect resolution, per-device comparison, reputation, certificates, content intent, brand correlation, domain intelligence — and only then, and only for entitled accounts, does anything AI-shaped happen. Even then the model is prompted as a corroborating voice and is structurally prevented from originating a verdict on its own.
The part worth arguing about is the gate. It would be easy to flag everything faintly suspicious and look vigilant. Instead 'dangerous' requires corroboration from more than one independent signal, and a clean shortener landing on a reputable domain passes. A security app that cries wolf is training its users to ignore it, which is a worse outcome than not existing.
Support & privacy
Next
Dogoos
Know how they feel