Trust boundaries

The model the rules are built on, and where a value stops being safe.

A trust boundary is a place where a value leaves the context that validated it. The rules do not look for dangerous functions; they look for values crossing those places.

The boundaries that matter here

In a Next.js App Router project, three crossings account for most of the risk:

Server Component → Client Component. Props you return from a Server Component are serialized into the React Flight payload and travel to the browser. Anything reachable from that object goes with them.

Server Action input. A Server Action is a public endpoint. Its parameters arrive from the network, whether or not the component that calls it is behind a login.

Server-rendered markup. A value interpolated into markup is in the HTML response, in the same place a value in props is.

Why the type system doesn't catch it

An object crossing into client props is well-typed either way. User is User whether it carries an email, a hashed password, or an internal risk score. Types describe shape. The question here is about provenance: where the value came from, and what has happened to it since.

Taint tracking

That is what the analysis actually does. It treats values from a source as tainted, follows them through assignments, calls, and returns, and reports when a tainted value reaches a sink (a boundary crossing) without passing through something that clears it.

• Source: where a value enters, typically a database read through your data layer.
• Sink: the crossing that fires the finding.
• Sanitizer: a step that clears taint. Recognized by name, so the ones your project uses have to be declared in antra.config.json.

Reading a finding

A finding is the sink, plus the path the value took to get there. The path is the interesting part: it is what tells you whether the crossing is a bug or a sanitizer the analysis did not know about.

antra trace app/account/page.tsx:42

Where this leaves false positives

A static analysis of provenance cannot know your intent. It can only know that a value from a source reached a sink without a recognized step in between. So the two ways to silence a finding are to add the missing sanitizer declaration, which fixes every instance at once, or to suppress this one with a reason.