Who am I ?

-Security researcher, publishing since 2005.

-Past research : vulnerabilities in BIOSes, Microsoft Bitlocker, Truecrypt, McAfee Endpoint (Defcon 2008), PMCMA debugger (Blackhat USA 2011), « Rakshasa » supply chain backdoor PoC (Blackhat 2012), 2 SAP notes (2013).

-Speaker/trainer at HITB, CCC, Ruxcon...

-Co-founder of the Hackito Ergo Sum and NoSuchCon research conferences (France).

Disclaimer : contains research

This was supposed to be a short research on finding/exploiting a few cool low level bugs in sandboxes.

It ended up leading to more questions than answers on my understanding of what the industry is doing in the AV/sandbox space.

If you have better understanding, I'd really like if you took the time to explain me

([email protected],+PGP).

Disclaimer (rephrased)

WTF is the AV industry doing ? Well, I'm not so sure I understand anymore …

But read the qemu source code, it's an awesome source of knowledge on system programming, compilers, binary translation and plenty other things :)

What's hot in the AV industry in

2013 ?

AV industry : 2013 trends

-Desktop AV is essentially a thing of the past

-Focus moves technologies hopefully able to « detect 0days »[1] like sandboxing.

=> The new cool thing is emulation and sandboxing.

[1] Don't laugh yet.

How it all started... (/story telling)

CVE-2013-0640

(Adobe Sandbox bypass)

Adobe Reader and Acrobat 9.x before 9.5.4, 10.x before 10.1.6, and 11.x before 11.0.02 allow remote attackers to execute arbitrary code or cause a denial of service (memory corruption) via a crafted PDF document, as exploited in the wild in February 2013.

CVE-2013-0640

(Adobe Sandbox bypass)

Their « analysis » :

Here is the sequence of the ROP shellcode: msvcr100!fsopen()

msvcr100!write()

mvvcr100!fclose()

kernel32!LoadLibraryA()

kernel32!Sleep()

Upon loading the malicious library, it will enter a long sleep and ensure that the thread has not crashed because the whole stack in the thread is already manipulated for creating a ROP chain.

Their « analysis » :

Here is the sequence of the ROP shellcode: msvcr100!fsopen()

msvcr100!write()

mvvcr100!fclose()

kernel32!LoadLibraryA()

kernel32!Sleep()

Upon loading the malicious library, it will enter a long sleep and ensure that the thread has not crashed because the whole stack in the thread is already manipulated for creating a ROP chain.

Their « analysis » :

Here is the sequence of the ROP shellcode: msvcr100!fsopen()

msvcr100!write()

mvvcr100!fclose()

kernel32!LoadLibraryA()

kernel32!Sleep()

Upon loading the malicious library, it will enter a long sleep and ensure that the thread has not crashed because the whole stack in the thread is already manipulated for creating a ROP chain.

=> In trivial english, this is called bullshitting. They clearly have no idea what the exploit is trying to do here.

What I believe really happens in this

case (wild guess)

Sleep 5 minutes to attempt bypass sanboxing detection :)

After all, it's a hardened exploit, found in the wild and the first of its kind to bypass Adobe sandboxing technology...

Limits of such technologies (imho)

-Good at finding artefacts (it's still « something »).

-Pretty bad at understanding what is actually happening inside the exploit.

That being said...

The raise of sandboxes...

The raise of sandboxes...

The raise of sandboxes...

The raise of sandboxes...

Note to self : I don't find quite reasonable to add to your corporate network something nobody really understands.

Note : lack of third party assessment

Note : lack of third party assessment

Note : lack of third party assessment

Room for problems

(research leads)

Room for problems part I : general design/architecture

What's the trend, perceived

objective like

Current « genious » idea :

-Correlate/share more data to create information asymetry.

-To do that, most (all I've seen allow it, at least in non default mode) solutions now allow a malware to connect back to the internet[*].

[*]Idea being to correlate DNS/binary checksum informations over « campaigns » of attacks in the time.

A few facts on this...

-The whole corporate strategy over the past 15 years has been to segregate LANS, DMZs and the internet.

-Now you give a temporary shell to the attacker at network perimeter (proxies, mail gateways, wherever such sandboxing solutions exist)...

-… and what happens to your DNS ? To your http proxy cache ?

The Katsuni-Kaminsky attack

(having it both ways)

-Attacker can run a malware inside a sandbox.

-Sandbox allows attacker to connect back to the internet.

-Corporate DNS server is used as a recursive DNS server.

-Attacker has it both ways and can synchronize arbitrary spoofed packets emission from both inside and outside the network.

=> That's gonna be very « safe » for sure...

Not to mention...

-What happens if the malware, from inside the sandbox manages to attack other networks (say crowdstrike !) ?

-What happens if the malware can send back a modified copy of itself to the same network (smtp?) for more analysis, and more sandbox cpu time ?

=> It's all about implementation details really.

Is this « wormable » (yet) ?

-FireEye claims to work with 30 % of the TOP 100 Companies. That makes it not so hard to find...

-Their strategy is to synchronize malware information sharing... (allow exploits to drop exes on the sandboxes... ?!?)

-They cover you « 360 », from mail gateways to http proxy file downloads, etc.

-Ok, so how exactly do you prevent one malware to get endless free execution time inside your different sandboxes around the world ?

Room for problems part II :

Turning lame bugs into sandbox

Oracles

The problem

-Many online malware scanning engines run qemu (+ some various instrumentation and automation custom software)

-User doesn't get to see anything from the scanning process.

-Now, what if an attacker has a qemu lame DoS or endless loop PoC ?

Turning lame bugs into sandbox

Oracles

(aka : hacking online malware analysis tools... hrm)

-Attacker wants to perform an arbitrary computation inside the sandbox whose output is a binary result (eg : RIP < 0x1234).

-Attacker produces a malware that performs this test, and then crashes (or enters endless loop) the sandbox, leading to an observable error (no malware reported, possibly error message if online service). Otherwise, the malware performs a « dodgy » action likely to be reported (download some page from the internet).

And therefore...

Extention of this attack to combine this with an online C&C in case the malware can gain infinite CPU time is left as an exercise to the reader.

And while we're at it...

Since the malware could very well embed SSL private certificates to connect back to the C&C, sniffing this trafic doesn't give the observer any clue about what the malware is actually sending back.

Room for problems part III :

Such bugs do happen...

What degrees do projects like Xen or qemu really have in terms of security ?

A fair question is the relative maturity of such technology, not when it comes to support legacy Oses or current Oses, but hostile malware trying to hurt them.

Exempli gratia : typical bugs reported

(complexity, software maturity : format strings/symlinks or complex overflows ?)

x86_emulate: MOVSXD must read source operand just once (Xen Unstable, 21/09/2013)

Ok, so you're at miscomputing EIP on basic instructions such as MOVSXD (used in every single binary, or about that)...

Note : this is totally exploitable imho btw.

Room for problems part IV : exploit hardening and orders of magnitude

Let's count

-An intel instruction is typically 3 bytes big

-Adobe reader 11.0 has dlls Mbytes big (28M AcroRd32.dll). Total mapping of exe/dll is 64 Mb on disk (including all sections).

-Assuming a uniform distribution, every 3-byte sequence has an equal probability of 1/16777216 to occur (mind CISC arch btw : no need to start with real instruction).

-Assuming 64Mb of executable mapping, that means each 3-byte sequence would appear 67108864/16777216 = 4 times.

=> Practically, if an attacker knew a few malicious instructions, they would find them in the mapping of an exploited binary (presence inside a proper usable ROP gadget sequence isn't a given, need better model).

Qemu internals

Binary translation in 1 slide

-Concept somewhat similar to UBQT (C. Cifuentes), takes a binary, translates it to IR. But then executes it on a virtual cpu instead of outputing a binary on a specific target.

-Virtual cpu has very few instructions (~40).

-Very generic, very portable, very fast.

-Many arch supported (ARM, MIPS, SPARC, PPC, …)

Qemu supports two modes

-System (full) virtualization

-Kernel emulation (wine/Windows, linux).

While many of current implementations are likely to be using the first flavour, we'll focus on the later one, which promises great speed enhancements

–so people are going to use just that imho- , and a much greater attack surface... ;)

Qemu architecture

Understanding binary translation &

memory sharing

Party time !!

Qemu regression tests...

Demos

Conclusion

Interresting technologies. Cool hacking tools, usefull for researchers.

There is room for massive security problems, the devil being in the details. Imho, maybe not ready for Enterprise grade deployment though (remember I have not seen most of those products !).

No such a thing as third party assessment afaict.

From a strict game theory pov, your best interrest is probably to

have others use that, but stay away from such technologies...

The Katsuni-Grothendieck theorem

applied to silver bullet 0day

sandboxing :

« You can't have it both ways ! »