| Permission | Why |
|---|---|
activeTab |
Temporary access to the tab the user clicked SiteCheck on |
scripting |
Inject the scanner into that tab |
storage |
Keep the last scan per tab in chrome.storage.session |
None of these produces a permission warning at install.
SiteCheck never declares host_permissions, optional_host_permissions, content_scripts
or externally_connectable. scripts/validate-build.mjs fails any build that does, or that
requests a permission outside the allowlist above. Widening the allowlist requires updating
this document first.
The end-to-end test build (dist-e2e/) adds http://127.0.0.1/* so Playwright can scan the
local fixture server. It is built to a separate folder, and the validator rejects it, so it
cannot be released by mistake.
- Nothing leaves the browser. SiteCheck makes no network requests and has no analytics.
- Scan results are kept in
chrome.storage.session: Chrome clears it when the browser closes, and SiteCheck removes a tab's entry when the tab closes. Page scripts and content scripts cannot read it. - Stored page URLs keep only origin and path. Query strings, fragments and embedded credentials, which can carry tokens, are dropped before storage.
- Each finding keeps at most 50 elements, with HTML snippets cut to 300 characters.
- Logs contain error codes only, never page content or URLs.
- Exports are created in the popup and saved through the browser's own download, with no
downloadspermission and no upload. They contain page content (HTML snippets, selectors) and the page URL without its query string, so treat exported files as you would the page.
The CSV protection's known trade-off: a cell that legitimately begins with - or + (such as
-5) is exported as '-5. Correctness of a spreadsheet formula is never worth the risk.
| Threat | Control | Status |
|---|---|---|
| Page content rendered as HTML (XSS in popup) | React text nodes only; ESLint bans dangerouslySetInnerHTML and innerHTML assignment; component test SC-UI-004 |
Active |
javascript: or other unsafe help links |
Only https: help URLs survive normalization and validation |
Active |
| Spoofed extension messages | Accepted only from SiteCheck's own extension pages (id and URL checked); other extensions and content scripts are ignored | Active |
| Malformed messages or scanner output | Runtime guards on every message and on the scanner result; stored state is re-validated when read | Active |
| Page-derived selectors used for highlighting | Passed to the page as serialized args and used only with querySelector; invalid ones report "not found" (SC-UI-005) |
Active |
| Help links opening attacker-controlled pages | https: only, opened with rel="noopener noreferrer" |
Active |
| Page tampering with the scanner | The scanner runs in the extension's isolated world, out of reach of page scripts | Active |
| Remote or dynamic code | Default MV3 CSP; build validator rejects CSP loosening; ESLint bans eval; axe-core is bundled, never fetched |
Active |
| Over-broad access | activeTab only, enforced by the build validator |
Active |
| Secrets in the package | Build validator rejects .pem, .crx and .env files |
Active |
| Compromised CI dependency | npm ci from the lockfile; exact dependency versions; actions pinned to commit SHAs; read-only token |
Active |
| Spreadsheet formula injection in CSV exports | Cells starting with =, +, -, @, tab or carriage return get a leading ' (OWASP); every field quoted (SC-SEC-002) |
Active |
| Unsafe characters in download file names | Host reduced to a-z 0-9 . -, capped at 60 characters; fixed sitecheck-… prefix (SC-UNIT-015) |
Active |
What was checked, and the evidence:
| Area | Finding | Evidence |
|---|---|---|
| Network | Scanning, highlighting and exporting make no network requests. | SC-SEC-003 records every request during a full session; a positive control proves the recorder works |
| axe-core stylesheet download | axe can re-fetch page stylesheets for a few CSS rules (preload). Now explicitly off. |
src/scanner/entry.ts; SC-SEC-003 |
| Dynamic code in our bundles | None. The popup's only fetch was Vite's module-preload polyfill, now disabled. |
validate-build.mjs fails any build whose popup or worker contains eval, new Function, fetch, XMLHttpRequest, WebSocket or sendBeacon |
| Dynamic code in axe-core | axe contains new Function (templates, plugin loading) in paths SiteCheck does not use. Chrome blocks string evaluation in the scanner's isolated world, so they could not run anyway. |
SC-SEC-004 gets EvalError from that world |
| Content Security Policy | Chrome's default also trusts http://localhost:* and http://127.0.0.1:*. SiteCheck now declares script-src 'self'; object-src 'self', and the validator requires it. |
SC-UNIT-001, SC-UNIT-002, SC-SEC-005 |
| Dependencies | npm audit: 0 vulnerabilities. Every version pinned exactly; fflate added for packaging only (build time, not shipped). |
package.json, package-lock.json |
| CI and release | Third-party actions pinned to commit SHAs; read-only token except the release job's contents: write; tag and version taken from environment variables, never interpolated into scripts. actionlint 1.7.12 reports no problems. |
.github/workflows/ |
| Packaging | The store zip is built only from a validated build, leaves out source maps, and is reproducible (identical bytes from the same commit) with a SHA-256 checksum. | SC-UNIT-016; CI's checksum step |
Run through this before every tag; record anything unusual in the release notes.
-
npm auditreports no known vulnerabilities, or each is assessed and noted. - The permissions table above still matches
manifest.config.ts; no new permission or host access without updating this document first. -
npm run validate:buildpasses (permissions, CSP, forbidden code, secrets). - The full CI pipeline passed on the commit being tagged.
- The regression suite's expected files are unchanged, or every change was reviewed.
- No secrets, keys or
.envfiles in the repository or package (git ls-files | grep -iE '\.(pem|crx|env)$'prints nothing). - Manual checks SC-MAN-001 and SC-MAN-002 done on the release build in desktop Chrome.
- The Chrome Web Store listing's privacy answers still match "Data handling" above.
- Chrome reports injection failures only as message text, which SiteCheck classifies by pattern. If Chrome rewords a message, the user sees the generic "scan failed" text rather than the specific one. Nothing unsafe follows from a misclassification.