$ cat writeup.md…
$ cat writeup.md…
sunshinectf2026
Task: web app leaks .git, MailHog, an encrypted 7z, and a hidden ChromaDB proxy storing password hints as 768-d embeddings. Solution: recover reverted git secrets, mine MailHog intel, abuse the proxy 403/502 oracle to dump embeddings, invert with vec2text, crack the SHA-256, decrypt the archive.
$ cat /etc/rate-limit
Rate limit reached (20 reads/hour per IP). Showing preview only — full content returns at the next hour roll-over.
English summary: A corporate-styled web app ("VecNet") for vector-database
security is hosted at https://vec.web.2026.sunshinectf.games/. Around it sits
a small infrastructure story: an exposed Git repository, a MailHog mail server,
a password-protected specs.7z archive containing the flag file, and a hidden
ChromaDB instance reachable only through a restrictive PHP reverse proxy on
port 8000. The password for the archive is never stored as text — only as a
768-dimensional text embedding plus a SHA-256 hash — so the final step is a
real embedding-inversion attack with vec2text.
The challenge is a chain of four loosely linked surfaces, each feeding the next:
index.html, fetch.php, and a
publicly listable /.git/ directory.VecNetDB (dimension 768)
where the password requirements are stored embedding-only
(document=null, metadata type=embedding_only,
embedding_fn=jxm/gtr__nq__32).jxm/gtr__nq__32 is the exact model pairing used by the vec2text library
(gtr-t5-base encoder, 768-d output — the author of vec2text trained these
inversion models). A 768-d embedding stored server-side with the plaintext
withheld is a textbook "embeddings leak the original text" setup, which the
MailHog emails even spell out by linking a dev.to article titled
"Vector Database Breaches: How Embeddings Expose Your Sensitive Data".
https://vec.web.2026.sunshinectf.games/.git/ returned an Apache directory
listing. Dump the whole repository object database:
...
$ grep --similar