PonyOS 0.4.99-mlp Privilege Escalation Vulnerability
John Cartwright <johnc@grok.org.uk>

Introduction
------------
Like countless others, I was pretty excited about PonyOS yesterday 
(April 1st 2013) and decided to give it a go.  After wasting a lot of 
time nyan'ing, I knew this was the future of desktop OSes.  However, I 
wondered how secure PonyOS really was.  So, I took a look at the 
source, which revealed that our ponies may be in danger of compromise!

All bugs tested against PonyOS 0.4.99-mlp from ponyos.org.

Userland Compromise
-------------------
Take a look at this snippet from login.c:

    int uid = checkUserPass(username, password);

    if (uid < 0) {
        fprintf(stdout, "\nLogin failed.\n");
        continue;
    }

    system("cat /etc/motd");

    pid_t pid = getpid();

    uint32_t f = fork();
    if (getpid() != pid) {
        /* TODO: Read appropriate shell from /etc/passwd */
        set_username();
        set_homedir();
        set_path();
        char * args[] = {
            "/bin/sh",
            NULL
        };
        syscall_setuid(uid);
        int i = execvp(args[0], args);

It seems that login runs 'cat' before dropping privileges. 

This is easy to exploit, given that the file permissions don't work.  
Just log in as 'local', and replace the 'cat' binary with another ELF 
- 'whoami' will do nicely for a PoC. Then log out, and back in again.

This causes your binary to run as uid 0. Exciting stuff!

Kernel Compromise
-----------------
Obviously userland exploits are boring and it was important that I 
find some kernel holes to play with.  Luckily PonyOS has quite a few 
for your enjoyment.

You can abuse syscall_fstat() to write the contents of the stat buf to 
an arbitrary kernel location if you so wish.  There are a few other 
similar bugs where pointers aren't sanitised, too.

    static int stat(int fd, uint32_t st) {
        if (fd >= (int)current_process->fds->length || fd < 0) {
            return -1;
        }
        fs_node_t * fn = current_process->fds->entries[fd];
        struct stat * f = (struct stat *)st;
        f->st_dev   = 0;
        f->st_ino   = fn->inode;

        ...

        f->st_mode  = fn->mask | flags;
        f->st_nlink = 0;
        f->st_uid   = fn->uid;
        f->st_gid   = fn->gid;
        f->st_rdev  = 0;
        f->st_size  = fn->length;

This is all well and good, but for today's 
silliness^h^h^h^h^h^h^h^h^himportant security audit I decided to 
exploit the ioctl handler found in tty.c:

    int pty_ioctl(pty_t * pty, int request, void * argp) {
        debug_print(WARNING, "Incoming IOCTL request %d", request);
        switch (request) {
            case TIOCSWINSZ:
                debug_print(WARNING, "Setting!");
                memcpy(&pty->size, argp, sizeof(struct winsize));
                /* TODO send sigwinch to fg_prog */
                return 0;
            case TIOCGWINSZ:
                memcpy(argp, &pty->size, sizeof(struct winsize));
                return 0;
            default:
                return -1; /* TODO EINV... something or other */
        }
        return -1;
    }

Printing WARNING to the console is fine, but the ponies won't get the 
message.  What we have here is pretty much an arbitrary read/write of 
kernel memory.

Want to read the value of 0x11223344 ?

    struct winsize ws;
    ioctl(0, TIOCSWINSZ, (void *)0x11223344);
    ioctl(0, TIOCGWINSZ, &ws);
    printf("%x %x %x %x\n", ws.ws_col, ws.ws_row, 
            ws.ws_xpixel, ws.ws_ypixel);
  
Want to zero the memory at that address?

    struct winsize ws;
    memset(&ws, '\0', sizeof(struct winsize));
    ioctl(0, TIOCSWINSZ, &ws);
    ioctl(0, TIOCGWINSZ, (void *)0x11223344);

Using these two primitives it is possible to dump out a large chunk of 
the kernel memory, find the process list, user_t, etc and patch it 
appropriately to change your uid.

There's a screenshot of an exploit in action at:
http://www.grok.org.uk/advisories/findus.jpg

However, the 'findus' code is not 'stable' enough to share right now.

Mitigating Factors
------------------
PonyOS doesn't come with a compiler, or any remote access, so it is 
quite difficult to exploit unless you build your exploit code into the 
OS image.

Having said that, there are some other bugs that could help you if you 
really wanted to attack PonyOS, given command line access.  For 
example, there is a perfectly good format string bug in the shell:

    for (int i = 0; i < shell_commands_len; ++i) {
        if (strstr(shell_commands[i], argv[0]) == shell_commands[i]) {
            list_insert(matches, shell_commands[i]);
            match = shell_commands[i];
        }
    }
    if (matches->length == 0) {
        list_free(matches);
        return;
    } else if (matches->length == 1) {
        for (int j = 0; j < strlen(context->buffer); ++j) {
            printf("\010 \010");
        }
    printf(match);

This can be triggered by going into /bin and creating an empty file 
called (for example) %08x%08x.  Perhaps you could use the excellent 
'bim' editor for this task.  Then, re-execute the shell, and use the 
tab-completion functionality to trigger the format string bug.  Just 
a few short pony-sized steps to go from there to injecting arbitrary 
code to exploit the system call issues.

Alternatively you might want to pass an invalid TERM to nyancat:

    char * nterm = getenv("TERM");
        if (nterm) {
            strcpy(term, nterm);
        }

Unfortunately the only way I could find to set this value was the 
shell's 'export' builtin, and long lines crash the shell (512 byte 
buffer...) so I haven't played with that bug yet, or indeed this other 
crash.

Conclusion
----------
There's so much to see and do! I don't think PonyOS will be replacing 
my other systems anytime soon, but it is an interesting project worthy 
of your attention.  I mean, it has ponies *and* massive security 
holes! What's not to like?

In all seriousness I accept the fact that the OS isn't meant to be 
secure in any way and I have essentially wasted 24 hours of my life 
horsing around with it.