`nosuid` and probably `nodev` should IMO be the default filesystem mount options. `/dev` is already a special devtmpfs and the initrd minimal /dev can just explicitly mount the initrd tmpfs rootfs with `dev` and `suid` if necessary.
Letting SUID binaries just "exist" anywhere is a stupendous security issue. What if you mount some external storage medium, how are you to verify that none of the SUID binaries on that block device are malicious.
Additionally, this exploit appears to only work if the user executing the SUID binary can also read the SUID binary. There's no reason for non-root users to have read on a SUID binary.
NixOS does this correctly. No SUID in the normal package installation directory `/nix/store` and no package leakage outside of that no `nosuid` can safety be used on all other mountpoints. The exception is just a single-purpose `/run/wrappers.$hash` directory that safety contains executable ONLY SUID wrappers.
While I hate suid as much as the next person, it's really not the problem here.
The bug that is being exploited gives you basically arbitrary page cache poisoning. At that point it's already game over. Patching a suid program is maybe the easiest way to get a root shell from that but far from the only.
The proof of concept exploit is just that. It is meant to demonstrate one attack vector only. There are many others. If your goal is to prevent the conceptual exploit only then there are many easier ways to accomplish that, such as blacklisting, that does not make you safer.
With this vulnerability you can manipulate the page cache. You could also manipulate ld.so to hook into arbitraty system calls, or set your uid to 0, or any of another dozen or so ways to elevate your privileges.
Mount points have nothing to do with this, even if is always a good idea to disallow suid in user writable areas and prevent reading suid files, but that's for other reasons. NixOS does nothing to fix this and is just as vulnerable as everyone else.
Without read permissions you cannot execute the binary, that would not make any sense.
To execute the binary it needs to be read from disk and loaded into memory.
In fact if you have read permissions but not executable permissions on a specific binary then you can still execute it by calling the linker directly /bin/ld.so.1 /path/to/binary (the linker will read and load the binary and then jump to the entry point without an exec() call)
> Without read permissions you cannot execute the binary
This is not correct, as when the binary is setuid-someone-else, you are not the one executing it; they are.
Removing world-readability from all setuid-root binaries on the system would be sufficient to kill the PoC script provided for this vulnerability. It would not be sufficient to prevent exploitation though; there are many ways to abuse the ability to write to files you have read access to in order to gain root, for example by using the vulnerability to alter the cached copy of a file in /etc/sudoers.d/, or overwrite /etc/passwd, or /etc/crontab, ... the list goes on.interesting but in that case no point in keeping the x bit either and suid binaries should just be 4700 ?
If they don't have world-execute permission, an access(2) check for executability would return negative, leading to things like shells not tab-completing it. The kernel would also deny attempting to execute it, as it is not executable for your fsuid.
You need execute access in order to launch it, but in order for it to run, the user it is running as (not you) needs read access; you don't.this ld.so magic will lose the suid bit
loader