io_uring takes a non-refcounted reference to the files_struct of the process that submitted a request (relying on ->flush() for being notified before the files_struct can go away). Unfortunately, unshare_fd(), which is used by bprm_execve() via unshare_files(), doesn't know about that, and assumes that if the files_struct's refcount is 1, it is okay to keep using the old files_struct. Therefore, an attacker can cause io_uring to fiddle around in the file descriptor table of a privileged process, and in particular steal file descriptors using IORING_OP_FILES_UPDATE.

References:

https://bugs.chromium.org/p/project-zero/issues/detail?id=2089
https://cdn.kernel.org/pub/linux/kernel/v5.x/ChangeLog-5.9.3
https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=0f2122045b946241a9e549c2a76cea54fa58a7ff