An X86 Design Flaw Allowing Universal Privilege Escalation

https://www.blackhat.com/us-15/briefings.html#the-memory-sinkhole-unleashing-an-x86-design-flaw-allowing-universal-privilege-escalation

by thomnottom • 11 years ago

167 66 11 years ago

blackhat.com

wmf• 11 years ago

I wonder if this is SMM or something even darker and deeper.

userbinator3• 11 years ago

SMM was my first thought too, although "gone unnoticed for 20 years" (1995) suggests something newer; SMM was introduced in 1990 with the 386SL, and this suggests P6-level.

Microcode seems about right - that was introduced in the P6. But, it's also signed (at least Intel ucode is - not sure about AMD) according to previous research... either way, this is going to be interesting.

mmastrac• 11 years ago

"directed against a uniquely vulnerable string of code running on every single system"

That seems to fit with the idea of microcode, but it's pretty vague. What else would be running on every system from the P6?

nly3• 11 years ago

If it is microcode it wouldn't be a complete disaster. Microcode updates are volatile and can ship with BIOS. Even when updated by Linux it happens early out of initrd.

dlitz• 11 years ago

Signed how, though? A crypto scheme that Intel came up with 20 years ago is probably not secure, and if they haven't updated it since then, well...

nly3• 11 years ago

"Memory sinkhole" seems be a pun on the "memory hole", which suggests it might be MMU related?

Zenst• 11 years ago

If you read "but 40 years of x86 evolution have left a labyrinth of forgotten backdoors into the ultra-privileged modes" then I'm more inclined to agree, though also wondering about V86 monitor after looking at http://www.scs.stanford.edu/05au-cs240c/lab/i386/s15_03.htm

Taniwha2• 11 years ago

I think you're right - SMM mode it is, we know the NSA has a bunch of SMM rootkits, I bet someonme found the vector the NSA uses

Moral_2• 11 years ago

to demonstrate how to jump malicious code from the paltry ring 0 into the deepest, darkest realms of the processor.

So you already have to be in ring 0. Pretty click baity title.

dmix• 11 years ago

"Privilege Escalation" implies you start from an unprivileged state does it not? That sentence might just be poorly structured.

geofft• 11 years ago

It means you go from some level of privilege (which could be none at all) to more privilege. For instance, "a WordPress author can get a shell on the server" would be a privilege escalation, even though being able to write blog posts is quite privileged on its own.

That said, I'm not particularly convinced that (unvirtualized) ring 0 to SMM is really a violation of a security policy... it's not like SMM really tries to confine what ring 0 can do, it just wants things like interrupt priority.

weland• 11 years ago

If you go from having your own ring 0 (as a guest in a hypervisor) to messing around with everyone's ring 0, it's very much a privilege escalation.

zerohp2• 11 years ago

VT-x allows guest operating systems to run in Ring 0. If this exploit works then a guest OS can exploit the hypervisor.

Moral_2• 11 years ago

Good call didn't tihnk about VT-x

tptacek2• 11 years ago

Maybe. VT hypervisors still mediate stuff done in VM kernels. The attack could still work on native OS kernels and still not be effective on VMs.

thomnottom• 11 years ago

Not intended as click bait. I saw it on twitter and was wondering just how serious a flaw people thought this could be. Or if the presenter might just be blowing smoke.

zobzu2• 11 years ago

well it still says "universal" but in fact, its universal if you are root/can insmod/etc.

it reminds me of the unlimited data plans that are limited after 5Gb.

Not sure what happened to language :)

exDM692• 11 years ago

So you already have to be in ring 0.

Yes, but the article implies that you can e.g. do bad things in a bootloader and then have the OS run while the exploit stays resident and undetected by the OS.

This definitely has previously undocumented exploit potential if you can force the target to e.g. boot from a malicious USB device.

edmccard• 11 years ago

if you can force the target to e.g. boot from a malicious USB device.

If you can control the boot media, aren't you already past the point where you need further exploits to control the machine?

frik2• 11 years ago

Intel CPUs support Ring0-3. But...

http://en.wikipedia.org/wiki/Protection_ring

Multics supported 8+ rings, OpenVMS 3 rings, OS/2 3 rings, Windows NT and Unix only 2 rings. Some hypervisor like XEN use afaik ring 1 (Intel VT-x "Vanderpool").

We need better ring support in modern operating systems and hypervisors (VM hosts).

"The x86-processors have four different modes divided into four different rings. Programs that run in Ring 0 can do anything with the system, and code that runs in Ring 3 should be able to fail at any time without impact to the rest of the computer system. Ring 1 and Ring 2 are rarely used, but could be configured with different levels of access."

j-pb• 11 years ago

Actually we should abandon the ring concept all together for more promising security models.

https://www.destroyallsoftware.com/talks/the-birth-and-death...

hga• 11 years ago

For better or worse, AMD dropped rings in x86-64.

praseodym• 11 years ago

Actually hypervisors (VT-x) run in Ring -1: http://en.wikipedia.org/wiki/Protection_ring#Hypervisor_mode

im3w1l• 11 years ago

What is more privileged than ring 0? Microcode?

zerohp2• 11 years ago

A hypervisor runs below ring 0 and system management mode runs below that.

ajross2• 11 years ago

... sort of.

"Ring 0" is a historical abstraction from the 80286 protected mode model. There was a two bit field associated with segment and gate descriptors that enforced privilege separation, so you couldn't load segment registers with data at a higher priviledge level, and were disallowed from making traps into higher levels except as specifically allowed (we call those "syscalls" today).

None of this stuff is used anymore. We have the kernel in ring zero and we have everything else.

A hypervisor is absracting the whole CPU, so the guests have their own rings, etc... SMM is likewise outside the ring model.

And of course we have all sorts of other priviledge abstractions in modern hardware: iommu's exist for this purpose of course (though with a different threat model), as does memory mapping handled by microcontrollers on the fabric of modern SoCs. The NX bit doesn't fit into "rings" but is clearly related technology, etc...

Basically we need to stop talking about 286 protected mode except when that's really what we mean. Frankly I have no idea what this attack means by "ring 0", but I'm guessing like everyone else this is an exploit in SMM code.

na85• 11 years ago

Where can a person read about the various modes and levels of abstraction without delving into x86 whitepapers and spec sheets?

Taniwha2• 11 years ago

SMM mode

smegel• 11 years ago

I remember a post a few years ago about running code by triggering page faults in the MMU. Basically they created a single instruction computer (which can be Turing complete) out of the MMU, hence allowing processing to occur with the MMU that is invisible to the CPU and everything else. I wonder if related.

http://en.wikipedia.org/wiki/One_instruction_set_computer

jbangert• 11 years ago

(author here) https://github.com/jbangert/trapcc

We did get a few hypervisor crashes and the Intel architecture has all sorts of subtle behaviour that is often not modelled properly. It would be good to see someone build on my work.

jhallenworld• 11 years ago

"deepest, darkest realms of the processor..." is it the ME? https://recon.cx/2014/slides/Recon%202014%20Skochinsky.pdf

userbinator3• 11 years ago

ME is in the chipset and it's much newer than 20 years.

nly3• 11 years ago

It's kind of scary that there's a micro-JVM on my chipset that allows new code to be loaded in to such a privileged system dynamically.

ymse• 11 years ago

Just like your mobile.. http://www.osnews.com/story/27416/The_second_operating_syste...

lmm• 11 years ago

Makes me think of this guy who managed to install linux on his hard drive: https://spritesmods.com/?art=hddhack&page=7

aji• 11 years ago

Is this about the IA-32 debugger? There's a Phrack article that describes something similar: http://phrack.org/issues/65/8.html

walterbell• 11 years ago

Earlier work from the ITL/Qubes team, including SMM attacks: http://invisiblethingslab.com/itl/Resources.html

xgbi• 11 years ago

Strangely, when I clicked the link my Macbook Air kernel-panicked. This is the first time it happened in maybe 2 years.. I though the website was actually using an exploit to kill all PCs kernels.

After a reboot all went fine, though.

0xFFC• 11 years ago

So guy's did I understand correctly ? They found a flaw in x86 so the guest can exploit host ? ( Maybe much restricted than what I said , but the main idea is correct ?)

vardump• 11 years ago

The worst case is a VM escape for hardware accelerated (VT-x/AMD-V). So does this exploit work under virtualized ring 0? That would be a disaster for many cloud providers and for virtualization in general.

Maybe this is something about controlled change of flow of execution in SMM mode?

tptacek2• 11 years ago

My guess is that if you have a universal hardware hypervisor escape, you write the abstract for that talk much differently.

tempodox• 11 years ago

Regardless of the content, white text on black background is VERY BAD for reading. Please make this text more readable.

bognition2• 11 years ago

Wow. Ok so given that this is a hardware flaw and not a software flaw is this kind of disclosure reasonable? In the case of software vulnerabilities, researchers typically inform vendors and allow them time to distribute a patch. Is there anything analogous here for hardware? If yes how does intel (or any hardware vendor) get ahead of this by distributing a fix?

JoeAltmaier• 11 years ago

If its an unfixable problem then there's an obligation to disclose. Else only the 'bad guys' know.

bognition2• 11 years ago

But doesn't this assume that the 'bad guys' have knowledge of the exploit. This feels like a faulty assumption. However after disclosure the bad guys will definitely have knowledge.

Also after disclosure what is the right course of action? We certainly can't require the replacement of all vulnerable hardware.

kllrnohj• 11 years ago

Ok so given that this is a hardware flaw and not a software flaw is this kind of disclosure reasonable?

Yes. Always, always yes. Even if it's unfixable, ignorance of there even being a problem is not better.

Is there anything analogous here for hardware?

Yes, actually. BIOS updates can include microcode updates and workarounds for hardware issues. Similarly if it takes a specific pattern to happen (which this almost certainly does as that's how these things work), that's going to be trivial for everyone's antivirus scanners to catch.

louwrentius• 11 years ago

This sounds very significant to me.

Basically this vulnerability affects multi-tennant systems where if you have a VM, you can then expoint this vulnerability to take control of a VM of somebody else.

This type of vulnerability, however rare, could be an argument why you should run systems with different security classifications on different physical hardware (different physical virtualisation clusters).

charonn02• 11 years ago

Am I missing something here? It's just a couple of sentences of some BS about negative rings.

scott_karana• 11 years ago

It's an exploit that'll be presented at the upcoming Blackhat conference.

charonn02• 11 years ago

Thanks, it makes much more sense sense now.

An X86 Design Flaw Allowing Universal Privilege Escalation

Related Stories