CATCH AND RELEASE: A NEW LOOK AT DETECTING AND MITIGATING HIGHLY OBFUSCATED EXPLOIT KITS BY MOHAMED SAHER AND AHMED GARHY
Jul 25, 2015
CATCH AND RELEASE: A NEW LOOK AT DETECTING AND MITIGATING HIGHLY OBFUSCATED EXPLOIT KITS
BY MOHAMED SAHER AND AHMED GARHY
AGENDA Our Intent
Rethinking Evasions
Domain of the Problem
Current Problem
Problem with Current Solutions
Solution #1 First Method
Solution #2 Second Method
OUR INTENT Is this function malicious?
function Translate(objects, offset, size) {
var length = 4;
for (var i = 0; i < size; i++) {
var r = rc.substr(0, length);
if(offset > 0) {
r = r.substr(offset) + r.substr(0, offset);
}
objects[i] = r.substr(0, r.length);
}
}
OUR INTENT Is this function malicious?
function Translate(objects, offset, size) {
var length = 4;
for (var i = 0; i < size; i++) {
var r = rc.substr(0, length);
if(offset > 0) {
r = r.substr(offset) + r.substr(0, offset);
}
objects[i] = r.substr(0, r.length);
}
} Without understanding the context on how a function is used, it is
very difficult to determine if it is malicious or not
OUR INTENT What about this script?
<script>
var a = '%25%33%43%69%66%72%61%6d%65 ...';
var b = unescape(unescape(a));
var spray = new Function(unescape(b));
</script>
OUR INTENT What about this script?
<script>
var a = '%25%33%43%69%66%72%61%6d%65 ...';
var b = unescape(unescape(a));
var spray = new Function(unescape(b));
</script> An “expert’s eye” can probably determine it looks suspicious.
The two are actually equal to each other
OUR INTENT What about this script?
<script>
var a = '%25%33%43%69%66%72%61%6d%65 ...';
var b = unescape(unescape(a));
var spray = new Function(unescape(b));
</script> An “expert’s eye” can probably determine it looks suspicious.
The two are actually equal to each other
Our intent is to allow an attack using the first example script, without depending on obfuscating like the second example script, and propose a more superior method for detecting both
RETHINKING EVASIONS Designing a new architecture
Use a message oriented architecture (MOA) to split the attack into disparate self contained messages – we refer to this as “units of work”
RETHINKING EVASIONS Designing a new architecture
Use a message oriented architecture (MOA) to split the attack into disparate self contained messages – we refer to this as “units of work”
This is a variation of the “script splitting” technique except a message exists within a local scope and is destroyed after it serves its purpose
RETHINKING EVASIONS Designing a new architecture
Use a message oriented architecture (MOA) to split the attack into disparate self contained messages – we refer to this as “units of work”
This is a variation of the “script splitting” technique except a message exists within a local scope and is destroyed after it serves its purpose
Does not require DOM manipulation to hide “magic strings”
RETHINKING EVASIONS Designing a new architecture
Use a message oriented architecture (MOA) to split the attack into disparate self contained messages – we refer to this as “units of work”
This is a variation of the “script splitting” technique except a message exists within a local scope and is destroyed after it serves its purpose
Does not require DOM manipulation to hide “magic strings” Avoid the “magic redirect IFRAME” that can be a trigger for some
analyzers
RETHINKING EVASIONS Designing a new architecture
Avoiding HTTP
An artifact that can be parsed or scanned for patterns, characteristics, and definitions does not exist
RETHINKING EVASIONS Designing a new architecture
Avoiding HTTP
An artifact that can be parsed or scanned for patterns, characteristics, and definitions does not exist
An alternative to loading JavaScript in “clear text”
RETHINKING EVASIONS Designing a new architecture
Avoiding HTTP
An artifact that can be parsed or scanned for patterns, characteristics, and definitions does not exist
An alternative to loading JavaScript in “clear text” Load one message at a time, forcing each message to be
analyzed independently – remember “units of work”
RETHINKING EVASIONS Designing a new architecture
Avoiding HTTP
An artifact that can be parsed or scanned for patterns, characteristics, and definitions does not exist
An alternative to loading JavaScript in “clear text” Load one message at a time, forcing each message to be
analyzed independently – remember “units of work” Web Sockets are a perfect candidate for both MOA and
bypassing HTTP from a web environment
RETHINKING EVASIONS Designing a new architecture
Avoiding HTTP
Avoiding client side state
Two components involved, client and server
Client
Listen
Invoke
RETHINKING EVASIONS Designing a new architecture
Avoiding HTTP
Avoiding client side state
Two components involved, client and server
Client
Listen
Invoke
Server
State
Send
RETHINKING EVASIONS Designing a new architecture
Avoiding HTTP
Avoiding client side state
Two components involved, client and server For each accepted connection from a client, server maintains a
state machine
RETHINKING EVASIONS Designing a new architecture
Avoiding HTTP
Avoiding client side state
Two components involved, client and server For each accepted connection from a client, server maintains a
state machine Messages are essentially commands and do not depend on each
other – remember “units of work”
RETHINKING EVASIONS Designing a new architecture
Avoiding HTTP
Avoiding client side state
Two components involved, client and server For each accepted connection from a client, server maintains a
state machine Messages are essentially commands and do not depend on each
other – remember “units of work” Client evaluates message, invokes message, and destroys it
RETHINKING EVASIONS Designing a new architecture
Avoiding HTTP
Avoiding client side state
Limit control flow and function call hierarchy
RETHINKING EVASIONS Designing a new architecture
Avoiding HTTP
Avoiding client side state
Limit control flow and function call hierarchy
Only client control flow is that of the client listening and invoking a message
RETHINKING EVASIONS Designing a new architecture
Avoiding HTTP
Avoiding client side state
Limit control flow and function call hierarchy
Only client control flow is that of the client listening and invoking a message
Order of messages not guaranteed by server. Server may send NOP messages as part of an attack to trick certain analyzers
RETHINKING EVASIONS Designing a new architecture
Avoiding HTTP
Avoiding client side state
Limit control flow and function call hierarchy
Only client control flow is that of the client listening and invoking a message
Order of messages not guaranteed by server. Server may send NOP messages as part of an attack to trick certain analyzers
“Monkey patch” functions dynamically evaluated in messages to trick certain analyzers
RETHINKING EVASIONS Designing a new architecture
Avoiding HTTP
Avoiding client side state
Limit control flow and function call hierarchy
Getting creative in transport format
RETHINKING EVASIONS Designing a new architecture
Avoiding HTTP
Avoiding client side state
Limit control flow and function call hierarchy
Getting creative in transport format
Web Sockets are simple TCP pipes, so data can be represented on the wire in an application specific way
RETHINKING EVASIONS Designing a new architecture
Avoiding HTTP
Avoiding client side state
Limit control flow and function call hierarchy
Getting creative in transport format
Web Sockets are simple TCP pipes, so data can be represented on the wire in an application specific way
No longer restricted to sending JavaScript in clear text
RETHINKING EVASIONS Designing a new architecture
Avoiding HTTP
Avoiding client side state
Limit control flow and function call hierarchy
Getting creative in transport format
Web Sockets are simple TCP pipes, so data can be represented on the wire in an application specific way
No longer restricted to sending JavaScript in clear text Create custom binary format
RETHINKING EVASIONS Designing a new architecture
Avoiding HTTP
Avoiding client side state
Limit control flow and function call hierarchy
Getting creative in transport format
Web Sockets are simple TCP pipes, so data can be represented on the wire in an application specific way
No longer restricted to sending JavaScript in clear text Create custom binary format Send message in binary on the wire
0100100001100101011011000110110001101111001000000100100001100001011011010110001001110101011100100110011100100001
RETHINKING EVASIONS Designing a new architecture
Avoiding HTTP
Avoiding client side state
Limit control flow and function call hierarchy
Getting creative in transport format
Web Sockets are simple TCP pipes, so data can be represented on the wire in an application specific way
No longer restricted to sending JavaScript in clear text Create custom binary format Send message in binary on the wire Simply looking at a binary message won't give hints about what its
contents are – is it an audio file, an image, even text?
RETHINKING EVASIONS Designing a new architecture
Avoiding HTTP
Avoiding client side state
Limit control flow and function call hierarchy
Getting creative in transport format
Web Sockets are simple TCP pipes, so data can be represented on the wire in an application specific way
No longer restricted to sending JavaScript in clear text Create custom binary format Send message in binary on the wire Simply looking at a binary message won't give hints about what its contents are
– is it an audio file, an image, even text? To even begin to understand a binary message, its format specification needs
to be known beforehand or else it is a very challenging problem in its own
RETHINKING EVASIONS Designing a new architecture
Avoiding HTTP
Avoiding client side state
Limit control flow and function call hierarchy
Getting creative in transport format
Confusing the Context
RETHINKING EVASIONS Designing a new architecture
Avoiding HTTP
Avoiding client side state
Limit control flow and function call hierarchy
Getting creative in transport format
Confusing the Context
Remember this function?function Translate(objects, offset, size) {
var length = 4;
for (var i = 0; i < size; i++) {
var r = rc.substr(0, length);
if(offset > 0) {
r = r.substr(offset) + r.substr(0, offset);
}
objects[i] = r.substr(0, r.length);
}
}
RETHINKING EVASIONS Designing a new architecture
Avoiding HTTP
Avoiding client side state
Limit control flow and function call hierarchy
Getting creative in transport format
Confusing the Context
Remember this function?function Translate(objects, offset, size) {
var length = 4;
for (var i = 0; i < size; i++) {
var r = rc.substr(0, length);
if(offset > 0) {
r = r.substr(offset) + r.substr(0, offset);
}
objects[i] = r.substr(0, r.length);
}
} Now that we get this from our binary format, we again ask the question, how do you determine if it is
malicious?
DOMAIN OF THE PROBLEM How can we define a malicious website?
How can we detect a malicious website?
How can we detect obfuscation?
DOMAIN OF THE PROBLEM How can we define a malicious website?
How can we detect a malicious website?
How can we detect obfuscation?
How can we identify obfuscation used for malicious purposes?
DOMAIN OF THE PROBLEM How can we define a malicious website?
How can we detect a malicious website?
How can we detect obfuscation?
How can we identify obfuscation used for malicious purposes?
How can we categorize what is malicious and what is not?
CURRENT PROBLEM Exploits delivered at some point relies on JavaScript
JavaScript is continuously getting obfuscated with more complexity
CURRENT PROBLEM Exploits delivered at some point relies on JavaScript
JavaScript is continuously getting obfuscated with more complexity
Current solutions are way behind in technology
PROBLEMS WITH CURRENT SOLUTIONS Relies heavily on invocative functions that are not a
concrete base to be malicious (fromCharCode, eval, unescape, etc.) and have plenty of legitimate use cases
PROBLEMS WITH CURRENT SOLUTIONS Relies heavily on invocative functions that are not a
concrete base to be malicious (fromCharCode, eval, unescape, etc.) and have plenty of legitimate use cases
DOM and CSS selectors
PROBLEMS WITH CURRENT SOLUTIONS Relies heavily on invocative functions that are not a
concrete base to be malicious (fromCharCode, eval, unescape, etc.) and have plenty of legitimate use cases
DOM and CSS selectors Client side proxies for client-server interaction
PROBLEMS WITH CURRENT SOLUTIONS Relies heavily on invocative functions that are not a
concrete base to be malicious (fromCharCode, eval, unescape, etc.) and have plenty of legitimate use cases
DOM and CSS selectors Client side proxies for client-server interaction Client side template engines
PROBLEMS WITH CURRENT SOLUTIONS Relies heavily on invocative functions that are not a
concrete base to be malicious (fromCharCode, eval, unescape, etc.) and have plenty of legitimate use cases
Limited sets of characteristics
PROBLEMS WITH CURRENT SOLUTIONS Relies heavily on invocative functions that are not a
concrete base to be malicious (fromCharCode, eval, unescape, etc.) and have plenty of legitimate use cases
Limited sets of characteristics
Probabilistic decisions is directly proportional with the characteristics extracted
TYPES OF APPROACHES Dynamic analysis of embedded JS
Static analysis of extracted JS (Method #1)
Static analysis of extracted JS (Method #2)
DYNAMIC ANALYSIS AdHoc Forwarding
Create a middle layer between the browser and the JS engine
Analyze the CFG of the scripts being executed
DYNAMIC ANALYSIS AdHoc Forwarding
Create a middle layer between the browser and the JS engine
Analyze the CFG of the scripts being executed Analyze a call hierarchy of functions order
DYNAMIC ANALYSIS AdHoc Forwarding
Create a middle layer between the browser and the JS engine
Analyze the CFG of the scripts being executed Analyze a call hierarchy of functions order Analyze certain combination of functions used including
known highly risky ones
DYNAMIC ANALYSIS AdHoc Forwarding
Browser Automation
Attach to IE process Use shdocvw.dll to automate COM callbacks
DYNAMIC ANALYSIS AdHoc Forwarding
Browser Automation
Attach to IE process Use shdocvw.dll to automate COM callbacks Capture events while they trigger and manipulate them
DYNAMIC ANALYSIS AdHoc Forwarding
Browser Automation
Attach to IE process Use shdocvw.dll to automate COM callbacks Capture events while they trigger and manipulate them Analyze in the same manner as AdHoc Forwarding
DYNAMIC ANALYSIS AdHoc Forwarding
Browser Automation
Browser In-Memory Injection
Inject JS in DOM to monitor events
DYNAMIC ANALYSIS AdHoc Forwarding
Browser Automation
Browser In-Memory Injection
Inject JS in DOM to monitor events Use a JS Debugger (FireBug or other)
STATIC ANALYSIS (METHOD 1) Analyze the script and categorize them based on certain
criteria
Web page encoding
STATIC ANALYSIS (METHOD 1) Analyze the script and categorize them based on certain
criteria
Web page encoding Detecting current language used and extracting features
STATIC ANALYSIS (METHOD 1) Analyze the script and categorize them based on certain
criteria
Web page encoding Detecting current language used and extracting features Check the WHOIS for the web page
STATIC ANALYSIS (METHOD 1) Analyze the script and categorize them based on certain
criteria
Web page encoding Detecting current language used and extracting features Check the WHOIS for the web page
Determine probabilistically to which category it belongs to
SHANNON’S ENTROPY Formula
We use Shannon’s Entropy to determine the entropy of the file only as a side-effect and not a main criteria to determine the decision whether it was malicious or not
NAÏVE BAYESIAN A machine-learning technique that can be used to predict
to which category a particular data case belongs
NAÏVE BAYESIAN A machine-learning technique that can be used to predict to
which category a particular data case belongs
Given the above formula’: An event A is INDEPENDENT from event B if the conditional probability is the same as the marginal probability
LAPLACIAN SMOOTHING To avoid having a 0 joint in any partial probability we use
the add-one smoothing technique
LAPLACIAN SMOOTHING To avoid having a 0 joint in any partial probability we use
the add-one smoothing technique.
Given an observation x = (x1, …, xd) from a multinomial distribution with N trials and parameter vector θ = (θ1, …, θd), a "smoothed" version of the data gives the estimator
where α > 0 is the smoothing parameter (α = 0 corresponds to no smoothing)
STATIC ANALYSIS (METHOD 2) How is JS executed/handled?
1. The code is scanned for all function(s) declaration. Each declaration is executed by creating a function object and a named reference to that function is created so that the function can be called from within a statement.
STATIC ANALYSIS (METHOD 2) How is JS executed/handled?
1. The code is scanned for all function(s) declaration. Each declaration is executed by creating a function object and a named reference to that function is created so that the function can be called from within a statement.
2. The statements are evaluated and executed by order as they appear on the page after fully loaded.
JS EXAMPLE #2
<script>
DoNothing();
</script>
<script>
function DoNothing() {
return;
}
</script>
This does not works
JS EXAMPLE #3
<script>
function DoNothing() {
return;
}
</script>
<script>
DoNothing();
</script>
This works
JS EXAMPLE #3
<script>
// assuming that DoNothing is not defined
DoNothing();
alert(1);
</script>
This does not works
JS EXAMPLE #3
<script>
// assuming that DoNothing is not defined
DoNothing();
</script>
<script>
alert(1);
</script>
This works
STATIC ANALYSIS (METHOD 2) Semantic analysis to focus on “what does this mean”
Optimizer-Compiler for JS which focuses on structure other than extracted invocative functions
OPTIMIZER-COMPILER The following describes the architecture of any ordinary
compiler and the current compiler as well
Lexer Parser Translator OptimizerTokens AST IR
OPTIMIZER-COMPILER At this phase the optimizer tries to optimize the JS input
based on optimization theories after the AST was generated and converted into an IR
Optimizer
Hidden Classes
OPTIMIZER-COMPILER At this phase the optimizer tries to optimize the JS input
based on optimization theories after the AST was generated and converted into an IR
Optimizer
Hidden Classes
Type Inference
OPTIMIZER-COMPILER At this phase the optimizer tries to optimize the JS input
based on optimization theories after the AST was generated and converted into an IR
Optimizer
Hidden Classes
Type Inference
Inline Caches
OPTIMIZER-COMPILER At this phase the optimizer tries to optimize the JS input
based on optimization theories after the AST was generated and converted into an IR
Optimizer
Hidden Classes
Type Inference
Inline Caches
Function Synthesis
OPTIMIZER-COMPILER At this phase the optimizer tries to optimize the JS input
based on optimization theories after the AST was generated and converted into an IR
Optimizer
Hidden Classes
Type Inference
Inline Caches
Function Synthesis
Inline Expansion
OPTIMIZER-COMPILER At this phase the optimizer tries to optimize the JS input
based on optimization theories after the AST was generated and converted into an IR
Optimizer
Hidden Classes
Type Inference
Inline Caches
Function Synthesis
Inline Expansion
Loop Invariant Code Motion
OPTIMIZER-COMPILER At this phase the optimizer tries to optimize the JS input
based on optimization theories after the AST was generated and converted into an IR
Optimizer
Hidden Classes
Type Inference
Inline Caches
Function Synthesis
Inline Expansion
Loop Invariant Code Motion
Constant Folding
OPTIMIZER-COMPILER At this phase the optimizer tries to optimize the JS input
based on optimization theories after the AST was generated and converted into an IR
Optimizer
Hidden Classes
Type Inference
Inline Caches
Function Synthesis
Inline Expansion
Loop Invariant Code Motion
Constant Folding
Copy Propagation
OPTIMIZER-COMPILER At this phase the optimizer tries to optimize the JS input
based on optimization theories after the AST was generated and converted into an IR
Optimizer
Hidden Classes
Type Inference
Inline Caches
Function Synthesis
Inline Expansion
Loop Invariant Code Motion
Constant Folding
Copy Propagation
Common Sub-Expression Elimination
OPTIMIZER-COMPILER At this phase the optimizer tries to optimize the JS input
based on optimization theories after the AST was generated and converted into an IR
Optimizer
Hidden Classes
Type Inference
Inline Caches
Function Synthesis
Inline Expansion
Loop Invariant Code Motion
Constant Folding
Copy Propagation
Common Sub-Expression Elimination
Dead Code Elimination