Please report security issues privately through GitHub's private vulnerability reporting for this repository. Do not open a public issue for anything that could be exploited before a fix ships.
You can expect an acknowledgement within 7 days. There is no bug bounty; this is a volunteer-maintained project. Reporters are credited in the changelog unless they ask not to be.
pullmd runs with one docker compose up. A report has to show the problem on
a running instance of the current release:
- the version tested (
GET /api/statusor the image tag) - the exact request(s) sent to pullmd and the response(s) received
- for SSRF claims, what the internal stand-in service logged
A reproduction against a re-implementation of the code, a static reading of a single file, or a proof of concept that was never run against pullmd itself does not meet this bar. Reports written by or with automated tooling are fine as long as the finding was verified on a real instance by a person who read this policy; reports that evidently were not are closed without reply, as are re-reports of the residuals listed below. No CVE is assigned for accepted residuals or for behaviour this document describes as intended.
In scope:
- The SSRF guard (
lib/ssrf.js) and every fetch path it protects, including redirects, sidecars and the MCP tools - Authentication, sessions and the OAuth 2.1 server
- Share links (
/s/:id), history scoping andDISABLE_PUBLIC_HISTORY - Input handling on
/api/htmland/api/file - The Docker images and compose files as shipped
Not security issues:
- Extraction quality, sites that block or rate-limit the fetcher
- Anything requiring a compromised host, container or
.env
The static fetch path (lib/ssrf.js) resolves a hostname, checks every
answer against the address policy and, since 3.18.2, opens the socket to
exactly the address it checked; a name that rebinds between check and
connect (DNS rebinding) is no longer followed there. This gap was disclosed
with the guard in PR #42
and was an accepted residual until then. Reddit short-link resolution
(redd.it, /s/ share links) used a plain fetch after the check until
3.19.2 and is pinned the same way since.
The Playwright sidecar applies the same address policy to every request Chromium makes (since 3.15.2, GHSA-26wr-prg5-vr77), but Chromium resolves names itself, so the DNS-rebinding residual remains on that path and is accepted. Operators who want a boundary that does not depend on the application can put the sidecar on its own bridge network with no other service on it; nothing in pullmd needs the sidecar to see the other containers. Please do not re-report it.
Text transcribed from Reddit post images (PULLMD_EMBEDDED_IMAGES, since 3.16.0)
is untrusted third-party content and can carry prompt injection. The output
labels it as machine-derived from the image; consumers should treat it like any
other fetched page text. This is inherent to describing arbitrary images and
not a vulnerability in pullmd.
Without sign-in (PULLMD_AUTH_MODE unset or disabled) and with public
history on (the default), every client that can reach the port can list
recent conversions including their share ids, and can delete cache entries.
This is the documented open mode, not a vulnerability. Operators who share an
instance should enable sign-in or set DISABLE_PUBLIC_HISTORY=true, which
also refuses anonymous deletes, or keep the port off public networks.
An upload sent with async=1 (POST /api/file) is converted in the background, and its result waits in memory under a random 128-bit job id until the client deletes it or the retention time passes (PULLMD_FILE_JOB_TTL_MINUTES, 60 minutes by default). It is never written to disk or to the cache, and a restart loses it. With sign-in enabled, a job is bound to the account that created it, and any other account gets the same 404 as for an unknown id. Without sign-in, the job id is the only credential, as with share links: anyone who has the id can read and delete the result.
Only the latest minor release receives fixes. Upgrade before reporting.