AI Error Message Humanizer — Rewrite Cryptic Errors with OpenRouter

AI Security Review

Evidence context: threat category none; evidence quality moderate.

This extension reads the error message at your cursor, sends it plus the surrounding code to a model on OpenRouter, and offers a rewritten message as a code action. Every finding in the scan follows from that one loop.

The network finding is NET-FETCH-extension/out/extension.js-1, a fetch call in extension/out/extension.js. That path is the compiled JavaScript build output, and the destination is openrouter.ai, the provider named in the extension's own description. A single HTTP call to a model API, inside a file whose only job is to call a model API, is the product working.

The 21 domain hits break down into strings the build carried along. openrouter.ai is the API. img.shields.io is a README badge. schemas.openxmlformats.org is an XML namespace identifier, a label inside a document format rather than a host the code contacts. errors.new is a Go error-handling site that shows up in prose about error messages. ciromaciel.click and by-ciromaciel.click are the only entries not tied to a third party, and both carry the publisher's own name, which fits a docs or landing page.

On credentials: there are no secret findings. Nothing in extension/out/extension.js shows reads of .env files, .git/config, SSH keys or cloud credential files, and no hit touches VS Code's secret storage. The extension needs exactly one credential, an OpenRouter API key, because the model calls are pay-per-token. The scan does not resolve how that key is stored or where it travels beyond openrouter.ai. Anyone installing it should assume the code context around an error reaches OpenRouter, because that is the transaction being sold.

The 14 code-smell hits and the single dependency hit are pattern matches against bundled npm code inside the minified build. Those rules fire on fetch calls, process.env references and crypto imports that any Node bundle contains. No malware signature matched, no obfuscation was flagged, and nothing points at tool poisoning or a modified IDE configuration.

The strongest counterargument is that this tool ships source context off the machine to a paid API and lists two .click domains, a cheap TLD that attracts abuse. That is fair to weigh. It does not change the conclusion: both .click domains match the author's name, the confirmed outbound call is the advertised one, and no code path reads secrets or writes outside the document you invoke the action on. Users who cannot send code to a third-party model provider shouldn't install it, and that is a question about OpenRouter's terms, which this scan does not answer.

Key Reasons

• NET-FETCH-extension/out/extension.js-1 resolves to extension/out/extension.js, the compiled build output, and targets openrouter.ai, the model provider named in the extension's own description.
• No secret findings: nothing in the scan shows reads of .env, .git/config, SSH keys, cloud credential files, or VS Code secret storage.
• Most of the 21 domain hits are non-endpoints: img.shields.io is a README badge, schemas.openxmlformats.org is an XML namespace identifier, errors.new is a Go error-handling site, openrouter.ai is the stated API.
• The 14 code-smell and single dependency findings match bundled npm code inside the minified extension/out/extension.js, which is build output from any non-trivial Node extension.
• No malware signature, no obfuscation, and no tool-poisoning findings; ciromaciel.click and by-ciromaciel.click carry the publisher's own name.

False Positive Considerations

• IoC extractor harvesting README badges and XML namespace URIs as network endpoints
• Code-smell rules firing on ordinary fetch, process.env and crypto patterns in bundled JavaScript
• Dependency finding generated by a package bundled into extension/out/extension.js
• Minified compiled output treated as suspicious structure rather than webpack/tsc build artifact

Reviewed 2026-10-18; recommended action: suppress false positive; model confidence 85%.