CVE-2015-5602 - Unauthorized Privilege Escalation

I'm a little surprised this got a CVE number to be honest; allowing users to edit files via some privileged mechanism when they may control some portion of the filesystem under consideration is always going to be dangerous.

sudo cannot actually prevent this -- for example, the patch for this issue http://www.sudo.ws/repos/sudo/rev/9636fd256325 (look for the sudo_edit.c hunk) just uses O_NOFOLLOW to try to mitigate this issue:

+#ifdef O_NOFOLLOW
+static int
+sudo_edit_open(const char *path, int oflags, mode_t mode, int sflags)
+{
+ if (!ISSET(sflags, CD_SUDOEDIT_FOLLOW))
+ oflags |= O_NOFOLLOW;
+ return open(path, oflags, mode);
+}
+#else

But O_NOFOLLOW only functions on the final component of a pathname, so you can still edit e.g. /etc/shadow if you create a symlink "ln -s /etc etc".

I'm pretty sure the sudo patch is more or less worthless; here's a far simpler program to test with:

| #include <sys/types.h>
| #include <sys/stat.h>
| #include <fcntl.h>
| #include <stdio.h>
|
| int main(int argc, char* argv[]) {
| int fd;
| int err = 0;
| int i;
|
| for (i=0; i<argc; i++) {
| fd = open(argv[i], O_RDONLY | O_NOFOLLOW);
| if (fd < 0) {
| fprintf(stderr, "open %s failed: %m\n", argv[i]);
| err++;
| } else {
| close(fd);
| }
| }
|
| return err;
| }

(Sorry for the pipes, launchpad collapsed all the spacing, making it illegible.)

$ make o_nofollow
cc o_nofollow.c -o o_nofollow
$ mkdir tests
$ cd tests
$ ln -s /etc etc
$ ln -s /etc/passwd passwd
$ ln -s /etc/shadow shadow
$ ../o_nofollow /etc/passwd /etc/shadow ./etc/passwd ./etc/shadow ./passwd ./shadow
open /etc/shadow failed: Permission denied
open ./etc/shadow failed: Permission denied
open ./passwd failed: Too many levels of symbolic links
open ./shadow failed: Too many levels of symbolic links
$

Note that opening ./etc/passwd succeeded here because ./etc is a symlink to /etc and O_NOFOLLOW does not prevent this.

The #else portion of the code may be fine, I haven't studied it extensively, so there may be some way to salvage the patch. But I suspect it's never going to be perfect.

Thanks

Ah, the demo program is still illegible only now with pipes. Sigh. I've attached the program here.

> The #else portion of the code may be fine, I haven't studied it extensively

I doubt this, it relies on comparing inode numbers and devices numbers returned by lstat() and fstat(). lstat() just like O_FOLLOW only considers the final component of the path. If it's a symlink, it returns data about the symlink otherwise it returns data about the file (even though it's accessed through symlinks).

Xenial now has 1.8.16, marking released.

The Precise Pangolin has reached end of life, so this bug will not be fixed for that release