Two easy bypasses of the "this link goes somewhere else" check:
- Any whitespace in the link text disabled it, so "Log in at
chat.lotusguild.org" → evil.example wasn't flagged. hostsFromText now
scans each word for host-shaped tokens; a link is flagged if any named
host is a different site. Bare file names (main.py, report.pdf) and
version numbers are skipped inside sentences so ordinary links aren't
flagged.
- Shared-hosting suffixes counted as one site (alice.github.io vs
evil.github.io, *.pages.dev, netlify.app, vercel.app, …).
registrableDomain now treats a curated list of them as public suffixes
(longest match, incl. s3.amazonaws.com). Curated rather than the full
PSL to stay inside the bundle budget.
Verified with unit tests and headless on real messages: the four bypass
cases are flagged; the honest sentence, file/version text and "click
here" are not.
Fixes#335
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_017PCb59uY3UKqDqt7B6hzd5
utils/linkSafety.ts compares the registrable domain the visible text claims
(when it looks like a URL/host) with the href's; a mismatch, or a punycode
(IDN) destination, renders the anchor as SuspiciousLink, whose click opens a
confirm — "It shows matrix.lotusguild.org but goes to evil.example." with the
full URL, Cancel / Open anyway (opens in a new tab with noopener). Honest links
are untouched: same registrable domain (youtube.com text over www.youtube.com,
bbc.co.uk over news.bbc.co.uk), plain-word text, mailto:, matrix.to and Lotus
permalinks, anchors with non-text children. Comparator unit-tested (incl. a
Cyrillic-а paypal homograph); verified headless that the phish and IDN
messages are flagged, the honest ones are not, the click shows the confirm and
does not navigate, Cancel keeps you put, Open anyway opens the real target.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PPmy3tPq869XDW4njjVaKA