Skip to content

Security: zootella/fuji

Security

security.md

Security, still open

  • Thumbnails and playback parse in the main process. The operating system's imaging library draws thumbnails there, and the table's pictures of the kinds only it reads, a TIFF on Windows and a Photoshop file on the Mac, behind the memory ceiling alone. Playback runs libmpv there too, where a Matroska file's font attachments reach FreeType through libass as its subtitle track starts, before any glyph is drawn and whether or not a subtitle uses them, read from mpv 0.41.0's sd_ass.c and libass 0.17.5's ass_fontselect.c. The sandbox process could take thumbnails and a video's still as it takes the probe. Playback draws into fuji's own window through mpv's render API on the Mac and Linux, so its narrower door is mpv's own options, against a file pulling in others from its folder and against its font attachments. With file as FFmpeg's only protocol, an HLS playlist named .m3u8 or .m3u can still open local files with media extensions, an ffconcat file can open files below its own folder, symlinks included, and a name holding %d, *, ? or { picks image2's file patterns, read from FFmpeg 9.0.2's hls.c, concatdec.c and img2dec.c.
  • Whether WebKitGTK sandboxes its content process on Linux is unchecked. On the Mac and Windows the web engine decodes the light table's pictures in a content process it sandboxes, and the Security page says so for those two alone.
  • A compromised page reaches whatever its commands reach. Code that took over the web engine's content process could call every command the page can, and the disk commands and process_open take any path, so it would reach the user's files and could start a program, one step from the main process. A registry of the folders the user has shown fuji, with every command that touches the disk refusing a path outside them, would narrow that.
  • On Windows the sandbox process can read what the user can, and reach the network. No Windows mechanism walls a process into one folder by itself. An AppContainer with no capabilities closes the network and every open by path, so the main process would hand it each file as a handle, duplicated with read access alone (DuplicateHandle), since the same access can come out larger. It needs read and execute granted on fuji's own executable and libraries in the per-user install folder, which Chrome's installer grants its own (configure_app_container_sandbox.cc), and whether an executable that imports user32 and WebView2 starts inside one is untested.
  • On Windows the sandbox process can end the main process. At the Untrusted level it can still open the process that started it for the right to end it, since ending sits in a process's execute rights, which the default policy of no read up and no write up leaves alone: a denial, not an escape. The thing to try is the main process labeling its own process object no-execute-up as it starts, after which the check's line for ending the process that started it should read refused.
  • The Mac's wall rests on an interface Apple does not support. Xcode 27's manual page, quoted on the OpenSSH list in August 2026, says a process using the named kSBXProfile profiles will be killed when built against SDK 27, and nothing found says custom profiles like fuji's are affected. Look again at each change of SDK.
  • Two bugs were open against the library fuji ships, as of 2026-10-15. CVE-2026-107678: an MP4 with tens of thousands of pssh boxes exhausts the stack through a recursive free and burns quadratic time in the MOV demuxer, its fix a pull request. CVE-2026-95127: signed overflow in libass 0.17.5 that wraps to garbage coordinates, its fix a pull request. Each fix reaches fuji through the recipe, one version and hash line per package, once it is released.
  • A pinned hash proves the bytes have not changed since they were pinned, not that they were the maintainers'. FFmpeg publishes a signature beside each release tarball, and some of the others do too, so checking it once when a hash line changes, and saying so in the library's record, would close that; mpv, SPIRV-Cross and shaderc come as archives GitHub generates from a tag, with nothing signed, and their record would say the tag and the hash are all there is.
  • The page has no top gate of its own. It has no unhandledrejection or error listener, so a rejection that escapes every caller's catch lands in the console nobody reads. One listener writing to the log would be the gate.
  • One call that never returns still freezes the light table. The table's queue moves only when the load at its head settles, so a load that never settles holds the table until fuji restarts; a timeout in the governors would end the wait. None of it can be tried on the disks fuji is developed on, so it needs a hang on demand: a share pulled away mid-read, a file system made to answer slowly, a FIFO named like a picture.
  • A file with no extension is never shown. The listing keeps only files whose extension fuji knows, so a JPEG saved as image is dropped before anything asks what it is. The tip scan names a file by its bytes, and a store of what it found would name such a file once.
  • A damaged file is refused whole. A truncated JPEG fails its decode and is counted. Whether to show the part that arrived, the way a browser does, is undecided.
  • A name that is not valid UTF-8 is lost in transit. disk_readdir hands names to the page as text, so a name Linux allows and an old collection may hold is replaced character by character, and can no longer be opened by what came back.

There aren't any published security advisories