New issue
Advanced search Search tips
Starred by 4 users

Issue metadata

Status: Fixed
Closed: Jun 2017

Sign in to add a comment

Issue 1223: MacOS/iOS userspace entitlement checking is racy

Reported by, Mar 23 2017 Project Member

Issue description

One way processes in userspace that offer mach services check whether they should perform an action on
behalf of a client from which they have received a message is by checking whether the sender possesses a certain entitlement.

These decisions are made using the audit token which is appended by the kernel to every received mach message.
The audit token contains amongst other things the senders uid, gid, ruid, guid, pid and pid generation number (p_idversion.)

The canonical way which userspace daemons check a message sender's entitlements is as follows:

  audit_token_t tok;
  xpc_connection_get_audit_token(conn, &tok);
  SecTaskRef sectask = SecTaskCreateWithAuditToken(kCFAllocatorDefault, tok);

  CFErrorRef err;
  CFTypeRef entitlement = SecTaskCopyValueForEntitlement(sectask, CFSTR(""), &err);

  /* continue and check that entitlement is non-NULL, is a CFBoolean and has the value CFBooleanTrue */

The problem is that SecTaskCreateWithAuditToken only uses the pid, not also the pid generation number
to build the SecTaskRef:

  SecTaskRef SecTaskCreateWithAuditToken(CFAllocatorRef allocator, audit_token_t token)
    SecTaskRef task;

    task = SecTaskCreateWithPID(allocator, audit_token_to_pid(token));

This leaves two avenues for a sender without an entitlement to talk to a service which requires it:

a) If the process can exec binaries then they can simply send the message then exec a system binary with that entitlement.
   This pid now maps to the entitlements of that new binary.

b) If the process can't exec a binary (it's in a sandbox for example) then exploitation is still possible if the processes has the ability to
   crash and force the restart of a binary with that entitlement (a common case, eg via an OOM or NULL pointer deref in a mach service.)
   The attacker process will have to crash and force the restart of a process with the entitlement a sufficient number of times to wrap
   the next free pid around such that when it sends the request to the target then forces the entitled process to crash it can crash itself and
   have its pid reused by the respawned entitled process.

Scenario b) is not so outlandish, such a setup could be achieved via a renderer bug with ability to gain code execution in new renderer processes
as they are created.

You would also not necessarily be restricted to just being able to send one mach message to the target service as there's no
constraint that a mach message's reply port has to point back to the sending process; you could for example stash a receive right with
another process or launchd so that you can still engage in a full bi-directional communication with the target service even
if the audit token was always checked.

The security implications of this depend on what the security guarantees of entitlements are. It's certainly the case that this enables
you to talk to a far greater range of services as many system services use entitlement checks to restrict their clients to a small number
of whitelisted binaries.

This may also open up access to privileged information which is protected by the entitlements.

This PoC just demonstrates that we can send an xpc message to a daemon which expects its clients to have the ""
entitlement and pass the check without having that entitlement.

We'll target which expects that only the cctool or sharingd binaries can talk to it.

use an lldb invocation like:

  sudo lldb -w -n corecaptured

then run this poc and set a breakpoint after the hasEntitlement function in the CoreCaptureDaemon library.

You'll notice that the check passes and our xpc message has been received and will now be processes by the daemon.

Obviously attaching the debugger like this artificially increases the race window but by for example sending many bogus large messages beforehand
we could ensure the target service has many messages in its mach port queue to make the race more winnable.

PoC tested on MacOS 10.12.3 (16D32)
6.0 KB View Download

Comment 1 by, Mar 23 2017

Project Member
Labels: Id-662011117 Reported-2017-Mar-23

Comment 2 by, Jun 9 2017

Project Member
Labels: -Restrict-View-Commit Fixed-2017-May-15
Status: Fixed (was: New)
Fixed in MacOS 10.12.5:
Fixed in iOS 10.3.2:

This issue was missing from the original published advisories but they were updated on Thursday June 8th 2017 to include this issue.

Comment 3 by, Jun 9 2017

Project Member
Labels: CVE-2017-7004

Sign in to add a comment