Summary
The remark42 image proxy fetches an arbitrary remote URL and re-serves the response from remark42's own origin. The download path decides whether the fetched resource is an image by looking only at the Content-Type header the remote server claims — it never inspects the actual bytes. The serving path then derives the response Content-Type by sniffing those bytes with http.DetectContentType.
An attacker hosts a URL that sets Content-Type to image/png but returns an HTML/JavaScript body:
- the download check sees
image/png → accepts it;
- the serve path sniffs the body → emits
Content-Type: text/html;
- the browser renders attacker HTML/JS as a document in remark42's origin.
Details
Downloader
backend/app/rest/proxy/image.go — downloadImage(), lines 189-206:
contentType := resp.Header.Get("Content-Type")
if !strings.HasPrefix(contentType, "image/") {
return nil, fmt.Errorf("invalid content type %s", contentType)
}
maxSize := 5 * 1024 * 1024 // 5MB default
if p.ImageService != nil && p.ImageService.MaxSize > 0 {
maxSize = p.ImageService.MaxSize
}
lr := io.LimitReader(resp.Body, int64(maxSize)+1)
imgData, err := io.ReadAll(lr)
if err != nil {
return nil, fmt.Errorf("unable to read image body: %w", err)
}
if len(imgData) > maxSize {
return nil, fmt.Errorf("image is too large")
}
return imgData, nil // <-- bytes never validated, returned as-isSend Content-Type: image/png and the check passes regardless of what the body actually contains.
Server
backend/app/rest/proxy/image.go — Handler(), line 131:
w.Header().Add("Content-Type", p.ImageService.ImgContentType(img))
_, err = io.Copy(w, bytes.NewReader(img))backend/app/store/image/image.go — ImgContentType(), lines 242-249:
func (s *Service) ImgContentType(img []byte) string {
contentType := http.DetectContentType(img)
if contentType == "application/octet-stream" {
return "image/*"
}
return contentType // <-- returns text/html for an HTML body
}PoC
self.send_response(200)
self.send_header("Content-Type", "image/png")
self.send_header("Content-Length", str(len(body)))
self.end_headers()
self.wfile.write(body) # body = <!DOCTYPE html><script>...</script>Then have the victim open https://<remark42-host>/api/v1/img?src=<base64(attacker-host)> top-level.
Impact
- The script can issue authenticated, same-origin API calls with
credentials: 'include' — the JWT cookie is sent automatically.
- The script can read the
XSRF-TOKEN cookie and re-send it as the X-XSRF-TOKEN header, defeating CSRF protection. The attacker acts as the victim: delete/edit their comments, change their settings, and — if the victim is admin — perform admin actions.
Triggering requires no remark42 account on the target instance; the attacker only needs to host the malicious upstream URL and deliver the proxy link to a victim by any means (email, DM, link on another site, etc.).
Fix
v1.16.0 adds layered defense to /api/v1/img and /api/v1/picture/{user}/{id}:
rest.SafeImgContentType validates sniffed body bytes against a strict allowlist (image/png, image/jpeg, image/gif, image/webp, image/bmp, image/x-icon). Non-image content returns 415 with no body echo. SVG is implicitly excluded.- Every response carries
Content-Security-Policy: default-src 'none'; sandbox; frame-ancestors 'none', X-Content-Type Options: nosniff, and Content-Disposition: inline; filename="image".
- The ETag is bumped to
"v2:<base64(src)>". Browsers that revalidate cached pre-fix responses get a fresh validated 200 instead of a 304 against the poisoned cached entry.
- The strict
default-src 'none'; sandbox CSP also applies to all /api/v1/* routes as defense-in-depth.
Residual exposure
Browser-local caches that already hold a pre-fix text/html response with Cache-Control: max-age=2592000 keep serving it from local store until the TTL expires or the cache is evicted under memory pressure. The ETag bump only reaches clients that revalidate during the cached lifetime. Operators running a CDN/edge cache in front of remark42 should purge /api/v1/img after deploying v1.16.0.
References
ildkh Reporter
Summary
The remark42 image proxy fetches an arbitrary remote URL and re-serves the response from remark42's own origin. The download path decides whether the fetched resource is an image by looking only at the
Content-Typeheader the remote server claims — it never inspects the actual bytes. The serving path then derives the responseContent-Typeby sniffing those bytes withhttp.DetectContentType.An attacker hosts a URL that sets
Content-Typetoimage/pngbut returns an HTML/JavaScript body:image/png→ accepts it;Content-Type: text/html;Details
Downloader
backend/app/rest/proxy/image.go—downloadImage(), lines 189-206:Send
Content-Type: image/pngand the check passes regardless of what the body actually contains.Server
backend/app/rest/proxy/image.go—Handler(), line 131:backend/app/store/image/image.go—ImgContentType(), lines 242-249:PoC
Then have the victim open
https://<remark42-host>/api/v1/img?src=<base64(attacker-host)>top-level.Impact
credentials: 'include'— the JWT cookie is sent automatically.XSRF-TOKENcookie and re-send it as theX-XSRF-TOKENheader, defeating CSRF protection. The attacker acts as the victim: delete/edit their comments, change their settings, and — if the victim is admin — perform admin actions.Triggering requires no remark42 account on the target instance; the attacker only needs to host the malicious upstream URL and deliver the proxy link to a victim by any means (email, DM, link on another site, etc.).
Fix
v1.16.0adds layered defense to/api/v1/imgand/api/v1/picture/{user}/{id}:rest.SafeImgContentTypevalidates sniffed body bytes against a strict allowlist (image/png,image/jpeg,image/gif,image/webp,image/bmp,image/x-icon). Non-image content returns415with no body echo. SVG is implicitly excluded.Content-Security-Policy: default-src 'none'; sandbox; frame-ancestors 'none',X-Content-Type Options: nosniff, andContent-Disposition: inline; filename="image"."v2:<base64(src)>". Browsers that revalidate cached pre-fix responses get a fresh validated200instead of a304against the poisoned cached entry.default-src 'none'; sandboxCSP also applies to all/api/v1/*routes as defense-in-depth.Residual exposure
Browser-local caches that already hold a pre-fix
text/htmlresponse withCache-Control: max-age=2592000keep serving it from local store until the TTL expires or the cache is evicted under memory pressure. The ETag bump only reaches clients that revalidate during the cached lifetime. Operators running a CDN/edge cache in front of remark42 should purge/api/v1/imgafter deployingv1.16.0.References