How a scan works

The domain-verification gate

This is the single most important thing to understand about your results, and the most common reason two scans look different.

Active probing does not run against a domain you have not proven you control. Until you verify, a scan runs the passive checks only — headers, cookies, DNS, JWTs, technology fingerprinting — and the report states plainly that active testing was skipped, rather than quietly returning a shorter list you might mistake for a clean bill of health.

Two ways to verify, either one is enough:

• A DNS TXT record at _secscan-challenge.yourdomain.com containing the token you are given. Proves control of the zone.
• A file at /.well-known/secscan-verification.txt containing the same token. Proves control of the web root.

The gate fails closed. If the verification lookup errors — a database blip, a DNS timeout — the answer is "not verified", not "assume yes". Failing open here would turn the scanner into precisely the thing the gate exists to prevent: an anonymous attack proxy that anyone can aim at anyone.

Why false positives are the hard part

A scanner that reports everything it might have found is worse than no scanner. Twenty speculative findings bury the one real one, and after the second false alarm nobody opens the report again. Three techniques do most of the work.

Catch-all fingerprinting

Any application whose router answers unknown paths with the app shell — every single-page app, and multi-tenant platforms like GitHub or npm — returns 200 for /.env and for anything else you ask for. Recording the response to a deliberately nonexistent path first, then suppressing every probe hit whose body matches it, removes that entire false-positive class. Some paths get further confirmation on top: /phpmyadmin requires a real login-form marker, and an OpenAPI hit requires a parseable schema rather than the word "swagger" on a page.

Baseline diffing

SQL injection is confirmed by fetching a benign baseline, then the same parameter with a trailing quote, and requiring a SQL error signature that was absent from the baseline. A page that always contains SQL-error-looking text is therefore never flagged.

Random canaries

Reflected XSS is tested with a per-request random token wrapped in HTML metacharacters. The finding stands only if the literal token reflects back with its angle brackets unencoded. An encoded reflection — which is to say, correct output encoding — produces no finding at all.

Safety rules for active probing

Deep probing asks real questions of your application. Every request is bounded by one rule: nothing is sent whose success would change customer data. In practice:

• Auth coverage uses GET and HEAD only. An unauthenticated POST sent to see whether it gets rejected performs the write when the answer is no.
• PUT is never sent, because many APIs treat it as an upsert.
• Write-verb authorisation is probed against an identifier nothing can exist at, so a successful DELETE destroys nothing while the status code still reveals whether the check ran.
• Mass assignment is read from the declared contract, not probed — sending a role field to see whether it sticks means granting somebody a role.
• The Supabase write probe uses an impossible column name. A schema rejection proves authorisation was bypassed without creating a row; if a permissive schema does create one, it is deleted immediately by its returned ID. No PATCH or DELETE is ever aimed at rows that already exist.
• SSRF callbacks point only at our own collector, never at an internal range or a cloud metadata address. The probe asks whether your server fetches a URL it was handed, and cannot be used to reach anything inside your network.
• A destructive-action guard keeps the crawler and the injection probes away from any URL that reads as delete, cancel, revoke or log out.

Scanning behind a login

Most of an application lives behind auth, and a scan that never signs in never sees it. Two modes: replay a session cookie or bearer token you paste in, or sign in through a login page with a headless browser.

Credentials are encrypted at rest, decrypted only inside the worker for the life of one scan, discarded when it ends, attached only to requests aimed at the scan target, and dropped on any redirect that leaves it. If the session expires mid-scan the scanner signs back in where it can — a form login renews itself, a pasted cookie cannot — and the report says so either way, because a scan that silently lost its session covers less than you asked for.

Supplying a second account unlocks the access-control comparison. The same record is requested three ways — as you, as somebody else, and as nobody — and the responses compared by token-overlap similarity rather than byte equality, because real pages carry CSRF tokens and timestamps that differ on every render. The anonymous leg is what keeps public pages out of the report: a product page returns the same bytes to everyone, which without that third request is indistinguishable from a leak.

What a black-box scan cannot tell you

Being honest about the boundary is part of the value. An external scan does not read your source, does not reason about your business logic, and cannot tell you whether a given user should have access to a given record — only whether two different accounts get served the same one. It does not replace a manual penetration test for an application handling regulated data.

What it does is find the entire class of problems that are mechanical, externally visible and embarrassing: the config file you deployed, the table you never put a policy on, the key that ended up in a bundle, the certificate nobody renewed. Those are worth catching automatically and repeatedly, which is exactly what a human tester is too expensive to do.