AI review

Besides its score, apm has a language model read new releases for code that works against the person installing them. A flag is a question for a person at apm, not a verdict.

How it works

  1. 1

    A model reads the release

    When a version is published on apm, a popular npm package ships a new one, a release adds an install script, or someone asks, a language model reads what changed since the previous version (the whole package for a first release). It looks for code that works against the person installing it.

  2. 2

    It reads twice

    A flag is checked with a second, separate read. Each finding has to cite a file and a line; a citation that does not hold up is dropped. If the two reads disagree, the version still waits for a person.

  3. 3

    A person decides

    A flagged version waits for someone at apm. They read the cited lines and either confirm the version is malicious, which blocks it everywhere, or dismiss the flag.

  4. 4

    Until then

    The package page says “Under review”. What the model found stays private until a person confirms it: models make mistakes, and a quoted line can contain a key the publisher shipped by accident. The package's own maintainers can see the findings for their versions.

The AI review cannot see dependencies, binaries or code a package downloads when it runs. It is one model's reading, separate from the score and not proof either way. What it reads and can't see