Fix new variant of the vulnerability in CVE-2023-35936.

Guilhem Moulin noticed that the fix to CVE-2023-35936 was incomplete.
An attacker could get around it by double-encoding the malicious
extension to create or override arbitrary files.

    $ echo '![](data://image/png;base64,cHJpbnQgImhlbGxvIgo=;.lua+%252f%252e%252e%252f%252e%252e%252fb%252elua)' >b.md
    $ .cabal/bin/pandoc b.md --extract-media=bar
    <p><img
    src="bar/2a0eaa89f43fada3e6c577beea4f2f8f53ab6a1d.lua+%2f%2e%2e%2f%2e%2e%2fb%2elua" /></p>
    $ cat b.lua
    print "hello"
    $ find bar
    bar/
    bar/2a0eaa89f43fada3e6c577beea4f2f8f53ab6a1d.lua+

This commit adds a test case for this more complex attack and fixes
the vulnerability.  (The fix is quite simple: if the URL-unescaped
filename or extension contains a '%', we just use the sha1 hash of the
contents as the canonical name, just as we do if the filename contains
'..'.)

src/Text/Pandoc/Class/IO.hs

@@ -224,6 +224,8 @@ writeMedia :: (PandocMonad m, MonadIO m)
            -> m ()
 writeMedia dir (fp, _mt, bs) = do
   -- we normalize to get proper path separators for the platform
+  -- we unescape URI encoding, but given how insertMedia
+  -- is written, we shouldn't have any % in a canonical media name...
   let fullpath = normalise $ dir </> unEscapeString fp
   liftIOError (createDirectoryIfMissing True) (takeDirectory fullpath)
   report $ Extracting (T.pack fullpath)

src/Text/Pandoc/MediaBag.hs

@@ -90,16 +90,17 @@ insertMedia fp mbMime contents (MediaBag mediamap) =
                        && Windows.isRelative fp''
                        && isNothing uri
                        && not (".." `isInfixOf` fp'')
+                       && '%' `notElem` fp''
                      then fp''
-                     else showDigest (sha1 contents) <> "." <> ext
+                     else showDigest (sha1 contents) <> ext
         fallback = case takeExtension fp'' of
                         ".gz" -> getMimeTypeDef $ dropExtension fp''
                         _     -> getMimeTypeDef fp''
         mt = fromMaybe fallback mbMime
         path = maybe fp'' (unEscapeString . uriPath) uri
         ext = case takeExtension path of
-                '.':e -> e
-                _ -> maybe "" T.unpack $ extensionFromMimeType mt
+                '.':e | '%' `notElem` e -> '.':e
+                _ -> maybe "" (\x -> '.':T.unpack x) $ extensionFromMimeType mt
 
 -- | Lookup a media item in a 'MediaBag', returning mime type and contents.
 lookupMedia :: FilePath

test/Tests/MediaBag.hs

@@ -20,7 +20,7 @@ tests = [
         let d = B.doc $
                   B.para (B.image "../../test/lalune.jpg" "" mempty) <>
                   B.para (B.image "moon.jpg" "" mempty) <>
-                  B.para (B.image "data://image/png;base64,cHJpbnQgImhlbGxvIgo=;.lua+%2f%2e%2e%2f%2e%2e%2fa%2elua" "" mempty) <>
+                  B.para (B.image "data:image/png;base64,cHJpbnQgImhlbGxvIgo=;.lua+%2f%2e%2e%2f%2e%2e%2fa%2elua" "" mempty) <>
                   B.para (B.image "data:image/gif;base64,R0lGODlhAQABAIAAAAAAAP///yH5BAEAAAAALAAAAAABAAEAAAIBRAA7" "" mempty)
         runIOorExplode $ do
           fillMediaBag d
@@ -35,4 +35,14 @@ tests = [
           (exists3 && not exists4)
         exists5 <- doesFileExist ("foo" </> "d5fceb6532643d0d84ffe09c40c481ecdf59e15a.gif")
         assertBool "data uri with gif is not properly decoded" exists5
+        -- double-encoded version:
+        let e = B.doc $
+                  B.para (B.image "data:image/png;base64,cHJpbnQgInB3bmVkIgo=;.lua+%252f%252e%252e%252f%252e%252e%252fb%252elua" "" mempty)
+        runIOorExplode $ do
+          fillMediaBag e
+          extractMedia "bar" e
+        exists6 <- doesFileExist ("bar" </> "772ceca21a2751863ec46cb23db0e7fc35b9cff8.png")
+        exists7 <- doesFileExist "b.lua"
+        assertBool "data uri with double-encoded malicious payload gets written outside of destination dir"
+          (exists6 && not exists7)
   ]