Triaging the exploitability of IE/EDGE crashes

****Introduction** **Introduction****

Both Internet Explorer (IE) and Edge have seen significant changes in order to help protect customers from security threats. This work has featured a number of mitigations that together have not only rendered classes of vulnerabilities not-exploitable, but also dramatically raised the cost for attackers to develop a working exploit.

Because of these changes, determining the exploitability of crashes has become increasingly complicated, as the effect of these mitigations must be taken into account during analysis. We have received a number of requests from the security community for clarification on how these mitigations affect exploitability. To ensure that only valid issues are submitted, we thought it may be useful to offer some guidance.

****Use after free mitigations** **Use after free mitigations****

Use-after-free (UAF) is a common type of vulnerability in modern object-orientated software. They are caused when an instance of an object is freed while a pointer to the object is still kept by the program. Since the object instance has been freed, this pointer is dangling, pointing to unmapped memory. Such a vulnerability is exploitable when the unmapped memory is controllable by an attacker, and will be used when the dangling pointer is later dereferenced by the program. We can split UAF vulnerabilities into 3 classes based upon where the dangling pointer is stored: the stack, heap, and the registers.

We have developed two primary mitigations to protect against UAFs:

• Memory Protector (MP)

MP is designed to protect objects against UAFs where the reference is stored on the stack, or in a register.

• MemGC

MemGC is a new replacement for MP, currently enabled on Edge and IE11. Protected objects are only freed when no references exist on the stack, heap or registers, offering complete coverage.

****Exploitability & Servicing** Exploitability & Servicing******MemGC [Edge & IE11] MemGC [Edge & IE11]

• We consider UAFs that are addressed by MemGC strongly mitigated, and will not issue a security update for them.
• The only exception for this are rare cases where zero writing the object leads to an exploitable state, although we have yet to see an occurrence of this.

Memory Protector [IE10 and below] Memory Protector [IE10 and below]

• We consider stack and register based UAFs strongly mitigated and will not issue a security update for them, except in the circumstances explained below.
• Heap reference based UAFs are not mitigated by MP, and so will still be addressed via a security update.

****Triaging crashes** Triaging crashes******Memory protector Memory protector

Memory protector (MP) is a mitigation first introduced in July 2014 initially for all supported versions of Internet Explorer, but now only applies to IE 10 and below. It is designed to mitigate a subset of use-after-free vulnerabilities, due to dangling pointers stored on the stack or the registers. At a high level, it works as follows:

1. When delete is called on an object instance, its contents is zero wrote, and it is placed in a queue. Once the queue has reached a threshold size, we then begin the process of seeing if it is safe to free each object instance in the queue.
2. To test to see if it is safe to free an object instance, we scan both the registers and all pointer aligned stack entries to see if there exists a pointer to the object. If no pointer is found then the object is freed, otherwise the object is kept in the queue.

Part (1) of the algorithm delays the potential freeing of the object to a later point in time, is controllable by an attacker, and as such is not considered a security mitigation.

To make it easier to determine the exploitability of these issues, MP has a mode called “Stress Mode”. Under this mode the delayed free component (1) of MP is disabled: stack/register scanning happens on every free, rather than when the queue has reached a threshold length. It can be enabled with the registry key:

HKLM:/Software/Microsoft/Internet Explorer/Main/Feature Control/FEATURE_MEMPROTECT_MODE/iexplore.exe DWORD 2

(note that this key, and “Stress Mode” are only applicable to MP, not MemGC).

Example crash Example crash

With the delayed free component of MP now disabled by forcing the object instance to be freed at the earliest possible instant, we can now concentrate on determining exploitability, based on Part (2), as shown by an illustrative example below:

](https://msdnshared.blob.core.windows.net/media/TNBlogsFS/prod.evol.blogs.technet.com/CommunityServer.Blogs.Components.WeblogFiles/00/00/00/61/47/2.png)

In this case, we have a use-after-free vulnerability causing a near-null dereference. Tracing backwards, we can see that the value of eax was set a few instructions previously:

](https://msdnshared.blob.core.windows.net/media/TNBlogsFS/prod.evol.blogs.technet.com/CommunityServer.Blogs.Components.WeblogFiles/00/00/00/61/47/3.png)

If we look at this object in memory, we see that has been zero wrote, and by checking the PageHeap End Magic we can see that this heap chunk is still allocated under Stress Mode:

](https://msdnshared.blob.core.windows.net/media/TNBlogsFS/prod.evol.blogs.technet.com/CommunityServer.Blogs.Components.WeblogFiles/00/00/00/61/47/4.png)

Now we need to see if there are any stack references to this object instance, starting at the call frame when delete was called. This can be completed using windbg scripting: for example, scanning for references to an object with base address stored in ebx with size 0x30:

](https://msdnshared.blob.core.windows.net/media/TNBlogsFS/prod.evol.blogs.technet.com/CommunityServer.Blogs.Components.WeblogFiles/00/00/00/61/47/5.png)

Checking stack reference locations with MP Checking stack reference locations with MP

In this case, we find a single reference to the object instance on the stack. With this information we must now check to see which call frame contains this reference.

Here, we show an example call stack at the point when the object is deleted:

](https://msdnshared.blob.core.windows.net/media/TNBlogsFS/prod.evol.blogs.technet.com/CommunityServer.Blogs.Components.WeblogFiles/00/00/00/61/47/6.png)

If there is a reference to an object instance on the stack or registers, then MP will never free the object instance. Thus, if between the point delete is first called in frame_2 until the point when we crash with a near null dereference in frame_5 there is always a stack reference, the object instance cannot be freed and reallocated/controlled by an attacker.

In this example, the reference we found by scanning the stack (at 0x1024ae9c) is stored in frame_8. Since this reference is present all of the time between the freeing point in frame_2 and the crashing point in frame_8, we consider this case as not-exploitable since it is strongly mitigated by MP.

Two other main situations can also occur:

1. If (for example) the stack reference was in frame_3 rather than frame_8, then there is a period between the freeing of the object and the crashing point when there are no stack references. This case may be exploitable since if the code path between these points can be slightly altered to force another call to delete, we will be left with an exploitable situation.
2. When running under stress mode, the crash may now occur on a freed block since the delayed free component is disabled (usually due to the reference being stored on the heap). Under this circumstance, the case would be generally exploitable.

MemGC MemGC

MemGC is a new replacement for MP, currently available in Edge and all supported versions of IE11, and mitigates use-after-free vulnerabilities in a similar fashion as MP. However, it also offers additional protection by scanning the heap for references to protected object types, as well as the stack and registers. MemGC will zero write upon free and will delay the actual free until garbage collection is triggered and no references to the freed object are found.

Just like MP, mitigated use-after-free vulnerabilities will most likely result in a near-null pointer dereferences or occasionally in no crash at all. If you suspect that a near-null pointer dereference is actually a mitigated use-after-free vulnerability you can verify this with the following steps:

• Find the position where the near-null value is read, determining the base pointer of the object:

](https://msdnshared.blob.core.windows.net/media/TNBlogsFS/prod.evol.blogs.technet.com/CommunityServer.Blogs.Components.WeblogFiles/00/00/00/61/47/7.png)

If we dump the object, we can see that it has been zero wrote as before:

](https://msdnshared.blob.core.windows.net/media/TNBlogsFS/prod.evol.blogs.technet.com/CommunityServer.Blogs.Components.WeblogFiles/00/00/00/61/47/8.png)

• Trace back and find the allocation call stack for this chunk, using the base pointer that was found in the first step. If the object is allocated with edgehtml!MemoryProtection::HeapAlloc() or edgehtml!MemoryProtection::HeapAllocClear() it means that the object is tracked by MemGC e.g.

](https://msdnshared.blob.core.windows.net/media/TNBlogsFS/prod.evol.blogs.technet.com/CommunityServer.Blogs.Components.WeblogFiles/00/00/00/61/47/9.png)

Similarly, when the object is freed, it will be via edgehtml!MemoryProtection::HeapFree() e.g.

](https://msdnshared.blob.core.windows.net/media/TNBlogsFS/prod.evol.blogs.technet.com/CommunityServer.Blogs.Components.WeblogFiles/00/00/00/61/47/10.png)

To double check that the issue is successfully mitigated, we can scan for references to the object on both the heap and stack.

For scanning the stack, we can use the same technique as described in the Memory Protector section. We can then use the same criteria as described above to determine exploitability; if there exists a stack reference between the freeing point and crashing point, we consider it strongly mitigated by MemGC.

When scanning the heap, we use a similar method, by first scanning the heap for references with values between the base pointer and basepointer+object_size of the object we are interested in. If any references are found, we then just need to check to see what objects they are associated with. If the object containing the reference is also tracked by MemGC (i.e. allocated via HeapAlloc() or HeapAllocClear()), then MemGC will not free the object we are interested in, so we consider it strongly mitigated by MemGC.

In this example, if we use the stack scanning command from above, we see that there is a reference on the stack preventing the object from being freed between the deletion and crashing points, making it successfully mitigated by MemGC.

](https://msdnshared.blob.core.windows.net/media/TNBlogsFS/prod.evol.blogs.technet.com/CommunityServer.Blogs.Components.WeblogFiles/00/00/00/61/47/11.png)

****Conclusions** **Conclusions****

In conclusion these new mitigations dramatically enhance the security by making sets of use-after-free vulnerabilities non-exploitable. When triaging issues in both IE & Edge, the behavior of these mitigations needs to be taken into account in order to determine the exploitability of these issues.

****Acknowledgments** **Acknowledgments****

We would like to thank the following people for their contribution to this post:

Chris Betz, Crispin Cowan, John Hazen, Gavin Thomas, Marek Zmyslowski, Matt Miller, Mechele Gruhn, Michael Plucinski, Nicolas Joly, Phil Cupp, Sermet Iskin, Shawn Richardson and Suha Can

–

Stephen Fleming & Richard van Eeden. MSRC Engineering, Vulnerabilities & Mitigations Team.