One leaked key reads every row of every table. We find it, then try to prove ourselves wrong.

No, and it never claims to be. Sentris reports reproducible findings with evidence, at a point in time. It is not a pentest, not a certification, and carries no warranty. If you need an attestation to close an enterprise deal, you need a pentest firm, not us.

No. The repo is never cloned. Files are read during the scan and discarded when it ends. What we keep is the findings, and any secret inside them is masked before it is written anywhere.

Only if you switch it on, per repository, and it is off for every repository until you do. When it is on, Sentris opens a pull request that adds exactly one new file: a migration closing findings it confirmed against your running system. It never pushes to your default branch, never replaces an existing line, never merges anything, and never opens a pull request for something it could not prove. Secrets are excluded on purpose: a pull request that deletes a leaked key creates false safety, because the key is still compromised and only you can rotate it. Leave the switch off and Sentris behaves exactly as it did before this existed.

You get a clean result that lists exactly what was checked and what is explicitly out of scope, so you know what the result covers. Then it keeps re-scanning on your plan's schedule (weekly on Starter, nightly on Pro and Team) and alerts you the first time something new appears. The value is the ongoing 'still clean', not only the first report.

Four of the deep scan's six stages have no model in them at all: the repo is partitioned by a table of rules, candidates are deduplicated deterministically, the coverage ledger is arithmetic, and suppression is a set operation. The model appears twice: once to read a component and once, in three independent lenses, to try to disprove what the first pass claimed. After that a finding has to quote your file character-for-character or it is dropped before it is stored. And with no model key configured, a deep scan returns exactly the deterministic rule checks rather than an error: the AI half can be absent without the product being broken.

Nothing is kept here to train on: the repo is never cloned and the files are discarded when the scan ends. During the scan itself the file contents do go to a model provider. That is what an AI pass is, and any tool that runs one and tells you otherwise is worth a second look. The current provider is named on the subprocessors page rather than in this answer, so it stays accurate when it changes. If you want none of it, leave the deep scan off: the rule checks are pattern matching on our own machines and never send your code anywhere.

Two ways. Rules classify by capability, so anon and publishable keys, which belong in your client, are never flagged. And anything code alone cannot prove is labelled potential rather than sold to you as a breach. Every detector has clean and vulnerable cases, and the clean repo must return zero findings before a repo rule ships. The repo and Deep Scan corpus is also the measurement: precision 1.000 and recall 1.000 across 38 planted findings in 11 repositories, published in full at /precision with the date, the commit and the command to reproduce it. Nobody else in this category publishes theirs.

Because the advisor reads your running database, and that is only one of the places these mistakes live. It cannot see the migration sitting in your branch before it ships, it cannot see the service_role key your client bundle is serving to every visitor, and it cannot see a route handler that takes an id and never checks who is asking. It also does not watch: it answers when you open it. Sentris covers all four surfaces and re-checks them on a schedule. If missing RLS is genuinely your only concern, use the advisor: it is free and it is good.

Every exposure the deterministic checks find, each with the exact file and line, masked evidence, and a copy-paste fix, plus re-scans on a schedule that alert you the first time something new appears. Scanning itself is free and needs no login: you see the count and the grade before you pay anything. What you are paying for is where they are and how to close them. Worth saying plainly: a URL-only scan reaches four of the checks, because the rest of them live in code a browser never sees. Connect the repo and a free scan runs all twenty rule checks instead of four, still without paying. The AI pass is the part that needs a subscription, because it spends tokens. No trial maze, no upsell ladder.

Yes, the whole thing, at /sample. It is an unedited scan of a deliberately vulnerable Supabase app we maintain, rendered by the same code that renders a customer's report: every finding, every file and line, the masked evidence and the SQL fix. We publish that repo rather than a customer's, and nothing on the page is trimmed for the screenshot.

One person, building on the same stack you do: Supabase, Next.js, shipped in evenings and weekends. Sentris is not a company with a security team behind it, and pretending otherwise would be the first thing worth distrusting about a security tool. What it is: focused checks that are published in full, run against fixed clean and vulnerable cases before they ship, and honest about the difference between what they proved and what they suspect.

Any time, from Stripe. Access follows your subscription status automatically, so there is nothing to email us about.