Skip to content

Security: ericovirgy/agentic-web-check

Security

SECURITY.md

Security policy

Supported versions

Version Supported
0.1.x yes
earlier no

Only the latest published minor receives fixes. Pre-releases (-beta, -rc) are not supported.

Reporting a vulnerability

Use GitHub's private vulnerability reporting: open the repository's Security tab and choose Report a vulnerability (https://github.com/ericovirgy/agentic-web-check/security/advisories/new). Do not open a public issue for a security problem.

Include the version (agentic-web-check --version), the command you ran, the target (a fixture under fixtures/sites/ if you can reproduce it there), and what happened. You should get an acknowledgement within a week. Fixes are published as a patch release with a CHANGELOG entry; you will be credited unless you ask not to be.

Scope

In scope: the scanner itself and everything it ships (src/, dist/, the GitHub Action under action/, the benchmark harness, the fixture server). Examples of things we want to hear about:

  • a page that can make the scanner execute code outside the browser sandbox, read files, or make network requests to hosts other than the target site and the configured LLM provider;
  • a way for page content to escape the task runner's safety guards (same-origin navigation, form-submit and consequential-action opt-ins);
  • API keys or other secrets leaking into reports, logs or artifacts;
  • path traversal in the fixture server or in --out handling.

Out of scope: the tool is not a security scanner for target sites. Its SAFETY dimension reports whether a site gives an agent what it needs to act safely (hidden instructions, unguarded consequential actions, unlabelled login boundaries, secret-looking tokens in page source). A pass is not a statement that the site is secure, and a finding about a third-party site is not a vulnerability in this project. Report those to the site owner.

Responsible use

  • Deterministic scans (no --tasks) load a handful of pages with an honest user agent (AgenticWebCheck/<version>), the same as any browser visit. Use judgement on sites that prohibit automated access.
  • Behavioural tasks drive a browser against the site. Only run them against sites you own or have permission to test. --allow-forms submits forms and --allow-consequential performs actions such as purchases or deletions; never enable them against a site that is not yours, and never put real personal data in a tasks file.
  • The benchmark's public dataset only lists sites that are built for or known to tolerate automated access. Keep it that way when proposing entries.

There aren't any published security advisories