This is a static page: HTML, CSS and three JavaScript files, served as files. There is no backend, no build step, no runtime dependency, no network call and no telemetry. Nothing you paste into it leaves your browser, and the only thing it persists is your last snippet, your two threshold values and your theme choice, in localStorage under the ai-code-detector: prefix.

That rules out most of what a security policy usually covers. What's left is worth reporting:

• Cross-site scripting. The app renders your input back to you — as line numbers, snippets and evidence rows. It escapes on the way out, but a bypass is a real finding.
• Prototype pollution or similar via a crafted snippet reaching the heuristics.
• A dependency issue in the dev tooling (playwright, http-server) that affects contributors running the audit or the tests.
• Anything that makes a network request. The masthead claims the page makes none, the fonts are vendored locally to keep that true, and a change that quietly breaks the claim is a privacy bug worth reporting as one.

Please report privately, not in a public issue:

Open a private security advisory via GitHub's "Report a vulnerability" flow. It's visible only to the maintainers until a fix is published.

Useful to include:

• what happens, and the smallest input that causes it;
• which browser and version;
• what you think an attacker gets out of it.

You should get an acknowledgement within a week. Since there is no server to patch and no released package to revoke, a fix ships as a commit to main and the deployed page picks it up on the next load — so the turnaround is usually short.

The deployed page and main are the same thing, and only the latest version is supported. Tagged releases are snapshots for reference; fixes land on main.

Out of scope, because they aren't defects:

• The verdict being wrong. The detector is a panel of heuristics with published weights, not a classifier with a guarantee. A false positive is a useful bug report (open an issue) but not a security one.
• Evading detection. Anyone can read the thirteen heuristics in detector.js and write code that dodges them. That's inherent to publishing the heuristics, which is the point of the project.