Permalink
Fetching contributors…
Cannot retrieve contributors at this time
103 lines (80 sloc) 4.42 KB

Problem

Integer overflow in pointer validation.

Discovered by

Ben Laurie <ben@links.org> using American Fuzzy Lop

Announced

2015-03-02

CVE

CVE-2015-2310

Impact

  • Remotely segfault a peer by sending it a malicious message.
  • Possible exfiltration of memory, depending on application behavior.

Fixed in

Details

The following text contains speculation about the exploitability of this bug. This is provided for informational purposes, but as such speculation is often shown to be wrong, you should not rely on the accuracy of this section for the safety of your service. Please update your library.

A specially-crafted pointer could escape bounds checking by triggering an integer overflow in the check. This causes the message to appear as if it contains an extremely long list (over 2^32 bytes), stretching far beyond the memory actually allocated to the message. If the application reads that list, it will likely segfault, but if it manages to avoid a segfault (e.g. because it has mapped a very large contiguous block of memory following the message, or because it only reads some parts of the list and not others), it could end up treating arbitrary parts of memory as input. If the application happens to pass that data back to the user in some way, this problem could lead to exfiltration of secrets.

The pointer is transitively read-only, therefore it is believed that this vulnerability on its own CANNOT lead to memory corruption nor code execution.

This vulnerability is NOT a Sandstorm sandbox breakout. A Sandstorm app's Cap'n Proto communications pass through a supervisor process which performs a deep copy of the structure. As the supervisor has a very small heap, this will always lead to a segfault, which has the effect of killing the app, but does not affect any other app or the system at large. If somehow the copy succeeds, the copied message will no longer contain an invalid pointer and so will not harm its eventual destination, and the supervisor itself has no secrets to steal. These mitigations are by design.

Preventative measures

In order to gain confidence that this is a one-off bug rather than endemic, and to help prevent new bugs from being added, we have taken / will take the following preventative measures going forward:

  1. A fuzz test of each pointer type has been added to the standard unit test suite. This test was confirmed to find the vulnerability in question.
  2. We will additionally add fuzz testing with American Fuzzy Lop to our extended test suite. AFL was used to find the original vulnerability. Our current tests with AFL show only one other (less-critical) vulnerability which will be reported separately (2015-03-02-2).
  3. In parallel, we will extend our use of template metaprogramming for compile-time unit analysis (kj::Quantity in kj/units.h) to also cover overflow detection (by tracking the maximum size of an integer value across arithmetic expressions and raising an error when it overflows). Preliminary work with this approach successfully detected the vulnerability reported here as well as one other vulnerability (2015-03-02-1). [See the blog post][4] for more details.
  4. We will continue to require that all tests (including the new fuzz test) run cleanly under Valgrind before each release.
  5. We will commission a professional security review before any 1.0 release. Until that time, we continue to recommend against using Cap'n Proto to interpret data from potentially-malicious sources.

I am pleased that measures 1, 2, and 3 all detected this bug, suggesting that they have a high probability of catching any similar bugs.