Thursday, August 9, 2018

BH18: Lowering the Bar: Deep Learning for Side Channel Analysis

Jasper van Woudenberg, Riscure

The old way of doing side channel analysis was to do leakage modeling to pull out keys from the signals. Started researching what happens if they use a neural network for the analysis.

They still need to attach the scopes and wires to the device, can't get robots to do that, yet. They do several runs and look for variations in signal/power usage to find leakages from the patterns (and divergence of the patterns).

Then we got a demo of some signal analysis - he made a mistake, and noted that is the problems with humans, we make mistakes.

Understanding the power consumption can give you the results of X (X-or of Input and Key), then if we know input - we can get the key! Still a lot of work to do.

In template analysis, you build models around various devices from power traces - then look for other devices using the same chipset, and then can start gathering input for analysis.

The researchers than looked at improving their processes with Convolutional Neural Networks (CNNS). THere is the input layer (size is equal to number of samples), the convolutional layer (feature extractor + encoding), then Dense Layers (classifiers) and finally the output later. Convolutional layers are able to detect the features independently of their positions.

There are a lot of visuals and live tracing, hard to capture here, but fascinating to watch :-)

Caveat - don't give too much input, make the network is too big = or the model cannot actually learn and will not be able classify new things.  (memorizes vs learning).  Need to verify this with validation recall. 

Deep learning can really help with side channel analysis and it scales well. It does require network fiddling, but it's not that hard. This automation will help put a dent into better securing embedded devices.


BH18: Legal Liability for IoT Cybersecurity Vulnerabilities

IJay Palansky, Partner, Armstrong Teasdale

IJay is not a cyber security expert, but he is a trial lawyer who handles complex commercial litigation, consumer protection,  and class actions - usually representing the defendant.

There is a difference between data breach and IoT vulns. They aren't handled the same. There is precedent on data breaches, but not really much on IoT devices. People have been radically underestimating the cost and volume of IoT lawsuits that are about to come. The conditions are going to be right for a wave of lawsuits.

Think about policy. The rules are changing. It is hard to predict how this will play out, so it's hard to say how IoT companies should protect themselves. IJay likes this quote from Jeff Motz - "What would make 'defense greater than offense'..?" (Motz? maybe Moss?)

People are trying to get the latest and greatest gadget out, to get the first to market advantage. Security slows this down. But if your'e not thinking about security devices up front, you are putting yourself at risk. If you get drawn into litigation or the media draws attention to it, you need to be able to answer to the media (or a judge) what you did to meet basic security requirements for that type of device. Think of avoiding the liability. Judges will look for who is the msot responsible.

It's estimated that there will be 20 Billion connected devices by 2020.

There are ridiculous items coming online all the time - like the water bottle that glows when you need to drink, the connected Moen shower to set temperature, and the worst the i.Con Smart Condom... oh boy.

These devices have potential to harm, from privacy issues to physical harm.  There can be ransomware, DDoS attacks, etc. These are reality - people are remotely hacking vehicles already.

Plaintiffs' lawyers are watching and wating, they want to make sure they can get soemthing out of it financially. They need to be able to prove harm and attribution (who to blame). Most importantly, the plaintiffs' lawyers don't understand this technology (and neither do the judges), or how the laws here work.

There is general agreement that the security of IoT devices is not where they should be. There will be lawsuits, once there are some, there will be more (those other attorneys will be watching).

This is not the first time that product liability or other law has had to address new technology, but the interconnectedness involved in IoT is unique. They need to show who's fault it was - could get multiple defendants, and they will be so busy showing what the other defendant did wrong - doing the plaintiffs' lawyer's job for them. :-)

There has been some enforcement by regulators, like the response to TRENDnet Webcam hack in Jan 2012, which resulted in a settlement in 2013.

Some lawyers will be looking for opportunities to take up these cases, to help build a name and reputation.

The Jeep hack was announced in 2015, then Chrysler recalled the vehicles. That's not where the story ends... there is a class action lawsuit moving forward still. (filed in 2016, but only approved yesterday to go forward). This is where things get interesting - nobody was hurt, but there was real potential of getting hurt.   People thought they were buying a safe car, and they were not. What is the value?

There is reputation loss, safety issues, and the cost of litigation that makes this all a problem. It's a burden and distraction on key employees that have to be deposed, find documents, etc.

The engineers and experts get stressed about saying something that will hurt their company, or thinking that they did something wrong that hurt someone. That is a cost.

IJay then walks us through law school in 10 minutes :-)

You need to understand the legal risks and assocaited costs, so when you are making decisions on the right level of security.

Damages vary by legal claim and the particular harm. Claims can be around things like negligence, fraud or fradulent omission, breach of warranty, strict product liability.  These are all state law claims, not federal, which means there will be variance.

Negligence means you have failed tot take "reasonable care" - often based on expert opinions.  Think of the Pinto - they had design defects.

Design defets could be around hardware or software, things like how passwords are handled.

Breach of warranty is an issue as well - there are implied warranties, like of merchantability (assumption product is safe and usable)  If you know you have an issue, and don't tell anyone - that's fraudulent omission.

Keep in mind that state statutes are dsigned to be cosnumer friendly, with really broad defintiions.

You need to minimally follow industry standards, but that may not be sufficient.

Think about security at all stages of your design, be informed and ask the right questions, be paranoid and allocate risk. Test and document the testing you did, save it while you do the work. It will hep protect you.  Be careful about words you use around your products, watch what you say in your advertisement and don't overstate what you do.

You should also get litigation insurance and make sure it covers IoT.

If it goes wrong - get a  good lawyer who knows this area. Investigate the cause, inclding discussions with engineers.

A wave of IoT hack and vuln litigation is coming - you need to be thinking about this now. Understand and use sound cybersecurity design and engineering principles.

BH18: WebAssembly: A New World of Native Exploits on the Browser

Justin Engler, Technical Director, NCC Group
Tyler Lukasiewicz, Security Consultant, NCC Group

WASM (WebAssembly) allows you to take code written elsewhere and run it in a browser.

Crypto minors and archive.org alike are starting to use web assembly.

Browsix is a project to implement POSIX interfaces in the browswer, and JsLinux has an entire OS in the browser. eWASM is a solution for ethereum contracts (an alternative to solidity). (and a bunch of other cool things)>

Remember when... Java Applets used to claim the same things (sandboxing, virtualization, code in browser)...

WebAssembly is a relatively small set of low-level instructions that are executed by browsers. It's a stack machine. You can push and pop things off the stack (to me the code looks a lot like lisp).  We do a couple of walkthroughs of sample code - they created a table of function pointers (egads! it's like networking kernel programming).

WASM in the browser - it can't do anything on it's own (can't read memory, write to screen, etc). If you want it to do anything, you need to import/export memory/functionality/etc. Memory can be shared across instances of Wasm.

Emscripten will help you create .wasm binaries rom other C/C++ code, incldues buit-in C libraries, etc.  Can also connect you to Java and JavaScript.

Old exploits in C work in WASM, like format strings and integer overflows. WASM has it's own integer types, different from C, different than JavaScript. You need to be careful sending integers across boundaries (overflow)..  Buffer overflows are an issue as well.  If you try to go past your linear memory, you get a JS error - it doesn't work well, it's pretty ugly.

You can now go from a BOF (Buffer Over Flow) to XSS. Emscripten's API allows devs to reference the DOM from C/C++. CHaracter arrays being written to the DOM create the possibilyt of DOM-based XSS and can use a user-tainted value to overwrite a safe value.  This type of attack likely won't be caught by any standard XSS scanners.  As JS has control of the WASM memory and tables, XSS should give us control of any running WASM.

And this even creates new exploits here! We can now have a function pointer overflow. Emscripten has functions that run arbitrary code (emscripten_rn_script). Can take advantage of that as lont as it's loaded. They discovered that function tables are constant - across compilations and even on different machines.

You don't necessarily to go after the XSS here, but could use functions written by the developers as long as it has the same signature as the real one.

They also showed a service-side RCE (Remote Code Execution). Showed code in browser starting a process on the server.

Many mitigations from C/C++ won't work on WASM. THey could use things like ASLR and could use some library hardening. Effective mitigations include control flow integrity and function definitions and indexing (prevents ROP-style gadgets).

WASM did cover these in their security warning, in a buried paragraph. It should be mroe obvious.

If you can avoid emscripten_run_script and friends, run the optimizer (removes automatically included functions that might have een useful for control flow attacks), use control flow integrity (but it may be slower) and you still have to fix your C bugs!

There is whitepaper out - Security CHasms of WASM

BH18: AI & ML in Cyber Security - Why Algorithms are Dangerous

Raffael Marty, VP Corporate Strategy ForcePoint

We don't truly have AI, yet. ALgorithms are getting smarter, but experts are more important. Understand your data and algorithms before you do anything with them. It's important to invest in experts that know security.

Raffael has been doing this (working in security) for a very long time, and then moved into big data. At Forcepoint, he's focusing on studying user behavior so that they can recognize when something bad is happening. ("The Human Point System")

Machine learning is an algorithmic way to describe data. In supervised case, we are giving the system a lot of training data. Unsuperfised, we give the system an optimization for it to solve.  For "Deep Learning" - it is a newer machine learning algorithm. It elminates the feature engineering step.  Data mining is a set of methods to explore data automatically.  And AI - "A program that doesn't simply classify or compute model parameters, but comes up with novel knowledge that a security analyst finds insightful" (not there, yet).

Computers are now better than people at playing chess and Go, they are even getting better at designing effective drugs and for making things like Siri smarter.

Machine learning is used in security, for things like detecting malware, spam detection, and finding pockets of bad IP addresses on the Internet in supervised cases, and more in unsupervised..

THere are several examples of AI failures in the field, like the Pentagon training AI to learn tanks (they used sunny pictures for "no tank" and cloudy with tanks, so the AI system assumed no tanks were in sunny weather... ooops!)

Algorithms make assumptions about the data, they assume the data is clean (often is not), make assumptions about distribution of data and don't deal with outliers.  The algorithms are too easy to use today - the proces is more important than the algorithm.  Algorithms do not take domain knowledge into account.  Defining meaningful and representative distance functions, for example.  Ports look like intergers and algorithms make bad assumptions here about "distance"

There is bias in the algorithms we are not aware of (example of translating "he is a nurse. she is a doctor" from English to Hungarian and back again... suddenly the genders are swapped! Now she is a nurse....)

Too often assumptions are made based on a single customer's data, or learning from an infected data set, or simply missing data.  Another example is an IDS that got confused by IKE traffic and classified it as a "UDP Bomb".

There are dangers with deep learning use. Do not use if there is not enough or no quality labelled data, look out for things like time zones along with timezones. You need to have well trained domain experts and data scientists to oversee the implementation, and understand what was actually learned.
Note - there are not a lot of individuals that understand security and data science, so make sure you build then a good, strong and cohesive team.

You need to look out for adversarial input - you can add a small amount of noise to an image, for example, that a human cannot see, but can trick a computer into thinking a picture of a panda is really a gibbon.

Deep learning - is it the solution to everything? Most security problems cannot be solved with deep learning (or supervised methods in general). We looked at a network graph - we might have lots of data, but not enough information or context nor labels - the dataset is actually no good.

Can unsupervised data save us?  Can we exploit the inherent structure within the adta to find anomalies and attacks?  First we have to clean the data, engineer distance functions, analyze the data, etc...

In one graphic, a destination port was misclassified as a source port (80!), and one bit of data had port 70000!  While it's obvious to those of us with network knowledge that the data is messed up, it's not to the data scientists that looked at the data. (with this network data, the data scientists found "attacks" at port 0).

Data science might classify port 443 as an "outlier" because it's "far" from port 80 - but to those of us who know, they are not "far" from each other technically.

Different algorithms struggle with clustered data, the shape of the data.  Even if you choose the "right" algorithm, you must understand the parameters

If you get all of those things right, then you still need to interpret the data. Are the clusters good or bad? What is anomalous?

There is another approach - probabilistic inference. Look at a Beysian Belief Networks. The first step is to build the graph, thinking about the objective and the observable behaviors. If the data is too complicated, may need to introduce "grouping nodes" and introduce the dependencies between the groups. After all the right steps, you still need to get expert opinions.

Need to make sure you start with defining your use-cases, bot by coosing an algorithm. ML is barely ever the solution to your problem. Use ensembles of algorithms and teach the algos to ask for input!  You want it to have expert input and not make assumptions!

Remember - History is not a predictor, but knowledge is"


BH18: Kernel Mode Threats and Practical Defenses

Joe Desimone, Gabriel Landau (Endgame)

Looking at kernel attacks, as it is a method to take over the entire machine and evade all security technology. Historically, Microsoft was vulnerable to malware - not prepared for those types of attacks, but they have made improvements over the year with things like PatchGuard and Driver Signature Enforcement. PatchGuard isn't perfect, attacks get through, but MS is constantly updating so the attacks don't work for long.

Both of these technologies are focused on 64-bit kernels, which is the growing norm today.

Attackers are now using bootkits, so Microsoft and Intel have come up with technology to counter (Secure Boot, Trusted Boot, Itnel Boot Guard, and Intel BIOS Guard).

All of those protections have changed the landscape. We don't see millions of kernel based botnets out there anymore.  But now people are signing their malware to look more legitimate and trick people to install.

DUQU 2.0 was a nation state attack, main payload used 0day in win32k.sys for kernel execution (CVE-2015-2360), it was spoofing process information to route maliious traffic on the internal network.

With the introduction of virtualization based security has also made the system more secure against things like Uroburos, Duqu2, DoublePUlsar.

The MS kernel has been greatly evolving over the last 10  years to greatly improved their mitigations. But, the problem is the adoption rate. There are still a lot of systems running Windows 7, which does not benefit from these new protections.

The speakers are on their orgs red team, so they are always looking for new ways to attack the system. They want to avoid detection and signature checks - their blue team is on the lookout for user mdoe priv escalation, so they wanted to be in the kernel. Looked at sample code from Winsock Kernel, found it was very effective (no beacons).

Did find a good attack, which means they needed to improve their own security.

Modification of kernel memory can significantly compromise the integrity of the system, so this is a major area of concern.

Need chip manufacturer to ship hardware with ROP detection enabled, otherwise this will always be a vector of attack. They did this by creating a surrogate thread, put it to sleep and though foudn the location of the stack and take advantage of it. (more details in the deck, the slides move pretty fast), but the interesting thing here is how much they can do by reusing existing code.

To project yourself, you should very carefully monitor driver load events. Look for low prevalence drivers and known-exploited drivers.  You need hypervisor protection policies, using white lists (which are hard to maintain) and leverage kernel drivers to WHQL. They have made a new tool to also  help to reduce the attack surface, available on their website today.

They wrote some code to generically detect function pointer hooks, locate the function pointers by walking relocation tables and leverage Endgame Marta.  They consider it a hit if it originally pointed to +X section in on-disk copy of driver, does not pont to a loaded driver in memory and points to executable memory.
 
ROP generates a lot of mispredictions, so need to protect this area as well (they could attack by scanning drivers to identify call/return sites, configure LBR to record CPLO near returns, etc)

The talk had lots of cool demos - can't really capture it here.

Windows platform security has gotten much better, but tehre are still kernel threats. You need to be using at least Windows 10 with SecureBoot and HVCI. at a minimum to protect yourself. Requite EV/WHQL within your organization

Wednesday, August 8, 2018

BH18: Don't @ Me: Hunting Twitter Bots at Scale

Jordan Wright, Olabode Anise, Duo Labs

Social media is a great way to have genuine conversations online, but the sphere is getting filled with bots, spam and attackers.

Not all bots on twitter are malicious - they could be giving us automated data on earthquakes, git updates, etc. So, their research was focused on finding bots and then figuring out if they were malicious.

The goal here is to build a classifier, one that could learn and adapt.

They wanted their research to be reproducible, so used the official Twitter APIs - though by doing so, they were rate limited. Because they were rate limited, they needed to be as efficient as possible. Fitting into that model, they were able to look up 8.6 million lookups per day.

Twitter's account ids started as sequential 32-bit unsigned integers, but the researchers started with random 5% sampling. The dataset has gaps - closed accounts, etc. Noticed accounts went up to very large numbers, and those accounts were up to 2016. But, Twitter changed to using "Snowflake IDs" - generated by workers, same format as other Twitter ids (tweets, etc).

The Snoflake ID is 63-bit, but starts with a timestamp (41-bits), then worker number (10 bits), then sequence (12 bits). It is very hard to guess these numbers. So, they used the streaming API with a random sample of public statuses (contains the full user object).

Now - they have a giant dataset :-)

Looked at last 200 tweets, accounts with more than 10 tweets, declared English and then they fetched the original tweets.  This data was too hard to get - could only do 1400 requests/day.

They took the approach of starting from known bots and discovering the bot nets they were attached to.

The data they have include attributes (how many tweets, are they followed, in lists, etc), looking at tweet content (lots of links?), and frequency of tweets.

They examined the entropy of the user name, was it fairly random? Probably a bot. Same for lots of numbers at the begining or end. Watchin for ratios of followers to following and the number of tweets.

They applied heuristics to the content - like number of hashtags in tweets, number of URLs (could be a bot or a news agency!), number of users @ replied.  On behavior - look at how long it takes to reply or retweet, and the unique set of users retweeted.  Genuine users would go queit for periods (like when sleeping).

Then we got a Data Science 101 primer :-)

This is where it gets complicated and statistics come into play, and the reminder that your model is only as good as your data. For example, if they trained with the crypto currency bots, they found 80% of the other spam bots. when reversed, they only caught about 50% of the crypto currency bots.


 Crypto currency give-a-way accounts are very problematic - they look legitimate and they will take your "deposit" and then you will lose your money.  They were hard to find, until they realized that there are accounts are out there that have many bots following them. Find those legitimate accounts, then you can find the bots.... also following like behaviors, used to map relatinships.  They found mesh and hub/spoke networks, but they were connected with likes.

They also discovered verified accounts that had been taken over, then they are modfiied to look like a more active account (like Elon Musk) that adds legitimacy to the crypto currency spam.

Very interesting research!



BH18: There Will Be Glitches: Extracting and Analyzing Automotive Firmware Efficiently

Alyssa Milburn & Niek Timmers, Riscure.

The standard approach for breaking into embedded systems: Understand target, Identify vulnerability, exploit vulnerability. Note - he also is referring to ECUs found in cars.

To understand the embedded system, need to understand the firmware. To do so - you need to get a hold of a car! Good source for cheep cars with common components - recalled Volkswagens :-)

Today's talk is targeting the instrument cluster - why? Because it has visual indicators you can see what is happening - it has blinking lights! :-)

Inside the instrument panel you will find the microcontroller, the EEPROM. display and the UART for debugging (but, it's been secured).  So, we have just inputs and outputs we don't understand. After much analysis, discovered most instrument panels talk UDS over the CAN bus. (ISO14229). This covers diagnostics, data transmission (read/write), security access check and loads more!

The team identified the read/write memory functions, but also discovered they were well protected.

Discovered that there are voltage boundaries, and if they go out of bounds they can stop the MCU. But... what if we do it for a very short amount of time? Will the chip keep running?

Had to get fault injection tooling - ChipWhisperer or Inspector FI - all available to the masses.

Fault injectors are great for breaking things. Once a glitch is introduced, nothing can be trusted. You can even change the executed instructions - opens a lot more doors! If you can modify instructions, you can also skip instructions!

They investigated adding a glitch to the security access check. Part of the check has a challenge, and if the expected response is received - access is granted. The team tried adding a glitch here, but were not successful, due to 10 minute timeout after 3 failed timeouts. As they are looking for something easy... moved on!

So, they mvoed on to glitching the ReadMemoryByAddress - no timeout here! They were successful on several different ECUs, which are designed around different MCUs.  Depending on the target, they could read N bytes from an arbitrary address. It took a few days, but were able to get the complete firmware in a few days.

There are parameters you can tweak for this glitch - delay, duration and voltage. Lots of pretty graphs followed.

It's hard to just do static analysis, as there is often generated code.

So, they wrote an emulator - allowed them to hook into a real CAN network, add debug stop points, and track execution more closely.

By using taint tracking, were able to find the CalculateKey function with the emulator.

There are new tools coming or electromagnetic fault injection - expensive right now, but getting cheaper.

ECU hardware still needs to be hardened - things like memory integrity and processing integrity. Unfortunately, these are currently being only designed for safety (not security).

There should be redundancy and the deisngers should be more paranoid. ECUs should not expose keys - need to leverage HSMs (hardened cryptographic engine). Highly recommend using asymmetric crypto - so the ECU only has a public key.

Do better :-)