Not actually crossing the airtight hatchway: Harmless out-of-bounds read that is never disclosed

Raymond Chen

A security vulnerability report arrived that went something like this:

By passing a specifically malformed payload, an attacker can trigger an out-of-bounds read. By this means, a remote attacker can cause the disclosure of sensitive information. An attacker can combine this with other vulnerabilities to achieve remote code execution.

The finder also included some reverse-compiled output¹ highlighting the point at which the out-of-bounds read occurred.

Anyway, it appears that the out-of-bounds read was discovered by using a memory debugging tool that does strict validations of every memory access. But consumers in the wild don’t run programs in such an environment.

When run on an actual consumer machine, the program uses the standard operating system heap manager, and the standard operating system heap manager does things like pad allocations to maintain alignment. Those extra bytes are technically off-limits, but they will always be there.

In this case, what happens is that the code allocates a block of memory, then reads past the end of it by a tiny amount, well within the heap padding, so it’s reading uninitialized heap memory. No denial of service is possible here because the heap padding saves you.

The next thing the code does is validate that the buffer is valid. This validation fails because the memory block is too small, and the operation fails. The value read from the uninitialized heap memory is not returned, so it is never disclosed to anybody.

Here’s a sketch. Assume that the checked_* functions reject the request if the operation fails.

struct ITEMSLIST
{
  uint32_t itemCount;
  ITEM items[ANYSIZE_ARRAY];
};

auto list = (ITEMSLIST*)checked_malloc(byteCount);
checked_read(list, byteCount);

auto requiredSize =
    checked_add(offsetof(ITEMSLIST, items) +
                checked_mult(sizeof(ITEM), header->itemCount));
checked_require(byteCount >= requiredSize);

... do stuff with the items ...

If the byteCount is less than sizeof(uint32_t), then the code under-allocates the list and tries to read the itemCount from it. Oh no, we are at risk of disclosing heap memory!

But then the code checks that the header size is large enough to hold the specified number of items, and seeing as the header size is not even large enough to hold the header, it certainly isn’t large enough to hold any items. So the request is rejected.

Note that the invalid itemCount never leaves the function. The value of itemCount is heap garbage, but whatever value it has will always fail the byteCount >= requiredSize test (assuming it manages to pass the checked_mult test), so the call will always be rejected. And the rogue value of itemCount is not exposed, so whatever garbage value happened to be there never escapes. What happens in parameter validation stays in parameter validation.

The finder jumped the gun: They found an out-of-bounds read but didn’t study it to see whether it was exploitable. They immediately concluded that there was information disclosure, and then tacked on a remote code execution for good measure.

What they found is a defect, but it has no security implications. It’s just a bug.

When informed that the issue as not exploitable and therefore has no security implications, they went ahead and issued a security bulletin anyway.

Six months later, the same organization found the same issue in a different component. We again told them that it was not exploitable and therefore has no security implications. The second time, they withdrew their plans to issue a bulletin.

So I’m not sure what changed over there, but at least they stopped issuing bogus bulletins for this category of issue.

Bonus chatter: This category of false alarm is quite common. People use various analysis tools to identify issues and immediately file a report without evaluating whether the issue actually is a vulnerability. They subscribe to the shotgun approach: File tons of potential issues, and let Microsoft figure out which ones are valid. Why do the extra work if you can externalize it!

¹ The reverse-compiled output has meaningless variable names like v1, v2 and v3, and object member accesses are expressed in the form (int*)((BYTE*)v40 + 0x20).

A note to people who send reverse-compiled output: Please include the original assembly language, and annotate that. Otherwise, we have to take your reverse-compiled output and try to re-compile it to assembly language in a way that matches the actual binary, and then re-reverse-compile it back to the original source code. These steps can be quite complicated because of compiler optimizations. (Also because people often fail to provide enough build number information to let us identify exactly which binary you are reverse-compiling, forcing us to keep trying all the different patched versions of the binary until we find a match or become exhausted.)

If you’re using IDA Pro’s Hex-Rays decompiler, you can right-click and say “Copy comments to disassembly.” That will take your comments in the reverse-compiled code and apply them to the corresponding lines of assembly.

7 comments

Discussion is closed. Login to edit/delete existing comments.

  • Lucius User 0

    I’d disagree that an out of bounds read that is never disclosed is harmless. While it is harmless in the binary-as-compiled, c compilers are becoming better and better at breaking the code in fun and exciting ways at even a slightest hint of undefined behavior (i am assuming it applies here because it looks like the original code is written in c±). In that case, since an out of bounds read is UB and can be assumed by the compiler to never happen, it can delete a bounds check that is located below the offending code.

    • smf 0

      I doubt it would just remove the bounds check, unless it could guarantee that the bounds check could never fail, which it obviously can fail.

      Any code paths where it can detect an actual out of bounds access could get optimized away entirely, because of undefined behavior. In which case there is no problem as code that is optimized away can’t have a security issue.

    • Ismo Salonen 0

      If the OOB read is from memory that is already mapped into your address space then technically there is no information disclosure but just a programming bug (in this case UB). If the read was outside your mapped address space and iff it succeeded then it would be information disclosure. Compiler does not always know if pointer is valid (imagine function which has char * argument, there is no other information passed than the pointer value).

      Of course all such bugs must be corrected, you never know if the pointer was just one byte past your address space and then you have unhandled exception and program crashes.

      • Julien Oster 0

        Absolutely not. A read from a region that is not mapped does not succeed, by definition. What would it return, since there is literally nothing there? An information disclosure happens whenever you return information that you did not intend to disclose. It’s your own address space with your own mappings, but you still can have secrets like, just as some examples, passwords, secret keys, or even just some random pointer that gives a good hint about your ASLR slide right next to your harmless and disclosable memory.

  • Kalle Niemitalo 0

    checked_mult could have a timing side channel.

  • Julien Oster 0

    Ok, I agree that it’s very likely not exploitable in this case, but I don’t actually think there is a guarantee that it is not, at least unless the heap padding is contractually guaranteed (which it likely is not).

    I know this is rather far fetched, but imagine that the out-of-bounds byte is not only outside any heap padding, but crosses into a mapped region in which reads have side effects. Like some MMIO device mapping, or (shudder) some user space code that implements its own faulting by catching the exception resulting from the unmapped read and mapping something on-demand, with side effects.

    I know you know that this is not the case here, because you know the heap layout and because you probably know that no other part of the program or library would map such side effect laden memory regions. But on some future OS version for example, it just might, even though this is admittedly one of the more unlikely bugs to ever turn exploitable. The bottom line is: Don’t assume too much.

  • Alexis Ryan 0

    I’d imagine concern over something like this because of heartbleed. the code may not expose anything but shows someone is writing dubious code

Feedback usabilla icon