fix(downloads): harden upload serving + ingest against stored-XSS and leaks (E7 review)

Deep-review remediation for the E7 upload path, all findings adversarially verified:

- Serve every library file with a Content-Type chosen from the ffprobe-verified media_kind
  + a strict container allow-list (storage.served_media_type), never guessed from the
  user-controlled file extension. Closes a stored-XSS vector where a media/HTML polyglot
  uploaded as "evil.html" was served inline as text/html on the login-free /watch surface.
  Applied to both the authenticated /file and the public watch_file endpoints.
- Classify uploaded still images (jpg/png/gif/webp/…) as media_kind='file' so they are
  force-downloaded and don't get a broken in-app video player.
- Roll back the on-disk side of an upload when the ingest transaction fails after place_file
  (unlink the placed file + its deterministic poster path), so a failed commit can't orphan
  bytes in the GC-skipped .uploads tree.
- Floor download_upload_max_bytes at min=1 so a stored 0 can't silently disable the per-file cap.
- Dedupe the ffprobe-JSON invocation into edit.run_ffprobe_json (shared by probe_codecs and
  upload._probe) and the upload extension into storage._clean_upload_ext; drop the now-unused
  parse_probe_streams.

Tests: public + authenticated force-download and the polyglot-not-served-as-HTML guard, the
real _probe classifier (image/cover-art/dims/duration), the ingest rollback/cleanup path, the
max_jobs branch, the cap-floor invariant, and run_ffprobe_json error handling.
This commit is contained in:
2026-08-29 21:52:02 +02:00
parent 104bd1e3a6
commit 614c5528ad
8 changed files with 475 additions and 128 deletions
+12 -6
View File
@@ -15,7 +15,7 @@ from sqlalchemy.orm import Session
from app.config import settings
from app.db import get_db
from app.downloads import links as linksmod
from app.downloads import mediakind, storage
from app.downloads import storage
from app.models import DownloadJob, DownloadLink, MediaAsset
from app.ratelimit import RateLimiter
from app.security import verify_password
@@ -122,11 +122,15 @@ def watch_file(token: str, g: str | None = None, db: Session = Depends(get_db)):
raise HTTPException(status_code=404, detail="This video is no longer available.")
filename = storage.download_filename(asset.title or "video", asset.container, path)
# A generic uploaded file (media_kind='file') is ALWAYS a force-download attachment with a
# neutral type — never served inline — so arbitrary uploaded HTML/SVG can't execute on this
# login-free public origin. Media follows the link's own inline/attachment choice so it can play
# The served Content-Type comes from the asset's ffprobe-verified media_kind + a container
# allow-list (storage.served_media_type) — NEVER guessed from the file's own extension. A generic
# 'file' (and any media whose container we don't positively recognise, e.g. a media/HTML polyglot
# named "evil.html") returns None → a neutral force-download attachment, so arbitrary uploaded
# HTML/SVG can't execute inline on this login-free public origin. A recognised media container is
# served with its real audio/video type, inline unless the link forces a download so it can play
# in the watch page. Both are range-aware (FileResponse → 206) for seeking/resume.
if asset.media_kind == mediakind.FILE:
media_type = storage.served_media_type(asset.media_kind, asset.container)
if media_type is None:
return FileResponse(
path,
filename=filename,
@@ -134,7 +138,9 @@ def watch_file(token: str, g: str | None = None, db: Session = Depends(get_db)):
content_disposition_type="attachment",
)
disposition = "attachment" if link.allow_download else "inline"
return FileResponse(path, filename=filename, content_disposition_type=disposition)
return FileResponse(
path, filename=filename, media_type=media_type, content_disposition_type=disposition
)
@router.get("/watch/{token}/poster.jpg")