Hallucinated and look-alike package names: checking a dependency before an agent installs it

Goal

Prevent an agent from installing a package whose name it produced from memory without confirming that the name belongs to the project it means.

Prerequisites

Registry access (PyPI, npm or another) from a sandbox; the ability to stop and ask before installation.

Steps

1. Treat every package name the model suggests as a claim. The paper "We Have a Package for You!" (Spracklen et al.) studied package hallucination in code generated by LLMs and found that models recommend packages that do not exist; an attacker who registers such a name receives the installs.
2. Resolve the name against the registry before installing. If it does not exist, do not look for "the closest match" — go back to the documentation of the library you actually need.
3. If it exists, confirm identity: the project's own documentation or repository names this exact package; the repository link in the registry metadata points back to that project.
4. Check for signs of a squatted or look-alike package: first release very recent, a single release, very few downloads compared with the name it resembles, a name one edit away from a popular package, a description copied from another project.
5. Inspect what runs at install time. npm runs lifecycle scripts such as preinstall and postinstall during npm install; installing with --ignore-scripts (npm's ignore-scripts setting) stops that, at the cost of breaking packages that genuinely need a build step. For Python, prefer wheels (--only-binary) so no build code runs.
6. Pin the confirmed version and hash in the lock file so later runs cannot drift to another artefact.
7. Report the check result with the package, version and evidence of identity in the change description.

Expected result

No package enters a project on the strength of a name the model generated; look-alikes are caught before their install scripts run.

Limits and test basis

A legitimate package can be compromised by its own maintainer account; identity checks do not detect that. Download counts can be inflated. The paper's rates depend on the models and prompts it tested and are not restated here.