mirror of https://github.com/dreamhunter2333/cloudflare_temp_email.git synced 2026-08-06 14:03:58 +08:00

Files

Josh Tsai e499211197 feat: add setting to disable auto-loading external images in emails (#1092 )

* feat: add setting to disable auto-loading external images in emails

Adds a privacy setting (default off) that blocks remote images in email
content until the user explicitly loads them per message. Blocked images
are replaced with a placeholder; a banner allows one-click loading.

Closes #1073

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* fix(frontend): block remote content with DOMPurify and an allowlist policy

Address review on the blocking logic. The first pass matched quoted
`<img src="http...">` with a regex, which left unquoted src, srcset,
`<source>`, CSS background-image, SVG `<image href>` and entity-encoded
schemes fetching as usual, and replaced only `src` on an element that also
carried `srcset` -- so the browser still had a remote candidate to prefer
while the UI claimed the image was blocked.

Two changes rather than a wider regex:

Sanitising is delegated to DOMPurify, which is already a dependency. The
hard part here is not enumerating attributes but surviving the parser: a
hand-written pass over a DOMParser tree still missed that `<noscript>` is
parsed as markup where scripting is off and as raw text where it is on, so a
`</noscript>` smuggled into an attribute value reopens the document at
render time and revives an `<img>` the cleaner never saw. Elements that
fetch by themselves or change how relative URLs resolve -- base, meta,
script, link, iframe, object, embed, noscript -- are dropped in this mode.
`<style>` is kept so layout survives, with its url(), image-set() and
@import references filtered.

URL classification is an allowlist. Asking "does this look remote?" means
enumerating every disguise -- backslash authorities, tab/newline/control
characters the URL parser strips, CSS escapes, schemes with no slashes --
and losing to the first one not thought of. Asking "can I prove this is
local?" fails closed instead: cid:, data:image/, blob: and relative paths
are kept, everything else is blocked. Relative paths are only safe because
`<base>` is removed, which is what stopped it re-pointing them at a tracker.

The blocked URL is discarded rather than parked in a data-* attribute, so
"the cleaned body contains no remote URL at all" is directly assertable;
restoring images re-renders from the untouched source.

Also: blob: is added to the allowed schemes -- DOMPurify's default list
omits it, and email-parser rewrites cid: attachments into blob: URLs, so
without it every inline image would be stripped along with the trackers.

The policy lives in its own module with its own tests (30 attack vectors,
7 preservation cases); email-parser.js goes back to MIME parsing only. The
per-mail override no longer initialises from the global setting, and the
banner reports the blocked count as the PR description promised.

Refs #1073

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: Claude Fable 5 <noreply@anthropic.com>

2026-07-29 14:28:15 +08:00