Unsupervised Machine Learning and JA3 for Enhanced Security
Unlock the true power of Darktrace's algorithms. Learn how JA3 enhances cybersecurity defenses with unique TLS/SSL fingerprints & unsupervised machine learning.
Darktrace cyber analysts are world-class experts in threat intelligence, threat hunting and incident response, and provide 24/7 SOC support to thousands of Darktrace customers around the globe. Inside the SOC is exclusively authored by these experts, providing analysis of cyber incidents and threat trends, based on real-world experience in the field.
Written by
Max Heinemeyer
Global Field CISO
Share
21
Jun 2018
Introducing JA3
JA3 is a methodology for fingerprinting Transport Layer Security applications. It was first posted on GitHub in June 2017 and is the work of Salesforce researchers John Althouse, Jeff Atkinson, and Josh Atkins. The JA3 TLS/SSL fingerprints created can overlap between applications but are still a great Indicator of Compromise (IoC). Fingerprinting is achieved by creating a hash of 5 decimal fields of the Client Hello message that is sent in the initial stages of an TLS/SSL session.
JA3 is an interesting approach to the increasing usage of encryption in networks. There is also a clear uptick in cyber-attacks using encrypted command and control (C2) channels – such as HTTPS – for malware communication.
The benefits of JA3 for enhancing rules-and-signatures security
These near-unique fingerprints can be used to enhance traditional cyber security approaches such as whitelisting, deny-listing, and searching for IoCs.
Let’s take the following JA3 hash for example: 3e860202fc555b939e83e7a7ab518c38. According to one of the public lists that maps JA3s to applications, this JA3 hash is associated with the ‘hola_svc’ application. This is the infamous Hola VPN solution that is non-compliant in most enterprise networks. On the other hand, the following hash is associated with the popular messenger software Slack: a5aa6e939e4770e3b8ac38ce414fd0d5. Traditional cyber security tools can use these hashes like traditional signatures to search for instances of them in data sets or trying to deny-list malicious ones.
While there is some merit to this approach, it comes with all the known limitations of rules-and-signatures defenses, such as the overlaps in signatures, the inability to detect unknown threats, as well as the added complexity of having to maintain a database of known signatures.
JA3 in Darktrace
Darktrace creates JA3 hashes for every TLS/SSL connection it encounters. This is incredibly powerful in a number of ways. First, the JA3 can add invaluable context to a threat hunt. Second, Darktrace can also be queried to see if a particular JA3 was encountered in the network, thus providing actionable intelligence during incident response if JA3 IoCs are known to the incident responders.
Things become much more interesting once we apply our unsupervised machine learning to JA3: Darktrace’s AI algorithms autonomously detect which JA3s are anomalous for the network as a whole and which JA3s are unusual for specific devices.
It basically tells a cyber security expert: This JA3 (3e860202fc555b939e83e7a7ab518c38) has never been seen in the network before and it is only used by one device. It indicates that an application, which is used by nobody else on the network, is initiating TLS/SSL connections. In our experience, this is most often the case for malware or non-compliant software. At this stage, we are observing anomalous behavior.
Darktrace’s AI combines these IoCs (Unusual Network JA3, Unusual Device JA3, …) with many other weak indicators to detect the earliest signs of an emerging threat, including previously unknown threats, without using rules or hard-coded thresholds.
Catching Red-Teams and domain fronting with JA3
The following is an example where Darktrace detected a Red-Team’s C2 communication by observing anomalous JA3 behavior.
The unsupervised machine learning algorithms identified a desktop device using a JA3 that was 100% unusual for the network connecting to an external domain using a Let’s Encrypt certificate, which, along with self-signed certificates, is often abused by malicious actors. As well as the JA3, the domain was also 100% rare for the network – nobody else visited it:
It turned out that a Red-Team had registered a domain that was very similar to the victim’s legitimate domain: www.companyname[.]com (legitimate domain) vs. www.companyname[.]online (malicious domain). This was intentionally done to avoid suspicion and human analysis. Over a 7-day period in a 2,000-device environment, this was the only time that Darktrace flagged unusual behavior of this kind.
As the C2 traffic was encrypted (therefore no intrusion detection was possible on the payload) and the domain was non-suspicious (no reputation-based deny-listing worked), this C2 had remained undetected by the rest of the security stack.
Combining unsupervised machine learning with JA3 is incredibly powerful for the detection of domain fronting. Domain fronting is a popular technique to circumvent censorship and to hide C2 traffic. While some infrastructure providers take action to prevent domain fronting on their end, it is still prevalent and actively used by attackers.
The only agreed-upon method within wide parts of the cyber-security community to detect domain fronting appears to be TLS/SSL inspection. This usually involved breaking up encrypted communication to inspect the clear-text payloads. While this works, it commonly involves additional infrastructure, network restructuring and comes with privacy issues – especially in the context of GDPR.
Unsupervised machine learning makes the detection of domain fronting without having to break up encrypted traffic possible by combining unusual JA3 detection with other anomalies such as beaconing. A good start for a domain fronting threat hunt? A device beaconing to an anomalous CDN with an unusual JA3 hash.
Conclusion
JA3 is not a silver bullet to pre-empt malware compromise. As a signature-based solution, it shares the same limitations of all other defenses that rely on pre-identified threats or deny-lists: having to play a constant game of catch-up with innovative attackers. However, as a novel means of identifying TLS/SSL applications, JA3 hashing can be leveraged as a powerful network behavioral indicator, an additional metric that can flag the use of unauthorized or risky software, or as a means of identifying emerging malware compromises in the initial stages of C2 communication. This is made possible through the power of unsupervised machine learning.
Darktrace cyber analysts are world-class experts in threat intelligence, threat hunting and incident response, and provide 24/7 SOC support to thousands of Darktrace customers around the globe. Inside the SOC is exclusively authored by these experts, providing analysis of cyber incidents and threat trends, based on real-world experience in the field.
The aim of this blog is to be an educational resource, documenting how an analyst can perform malware analysis techniques such as unpacking. This blog will demonstrate the malware analysis process against well-known malware, in this case SnappyBee.
SnappyBee (also known as Deed RAT) is a modular backdoor that has been previously attributed to China-linked cyber espionage group Salt Typhoon, also known as Earth Estries [1] [2]. The malware was first publicly documented by TrendMicro in November 2024 as part of their investigation into long running campaigns targeting various industries and governments by China-linked threat groups.
In these campaigns, SnappyBee is deployed post-compromise, after the attacker has already obtained access to a customer's system, and is used to establish long-term persistence as well as deploying further malware such as Cobalt Strike and the Demodex rootkit.
To decrease the chance of detection, SnappyBee uses a custom packing routine. Packing is a common technique used by malware to obscure its true payload by hiding it and then stealthily loading and executing it at runtime. This hinders analysis and helps the malware evade detection, especially during static analysis by both human analysts and anti-malware services.
This blog is a practical guide on how an analyst can unpack and analyze SnappyBee, while also learning the necessary skills to triage other malware samples from advanced threat groups.
First principles
Packing is not a new technique, and threat actors have generally converged on a standard approach. Packed binaries typically feature two main components: the packed data and an unpacking stub, also called a loader, to unpack and run the data.
Typically, malware developers insert a large blob of unreadable data inside an executable, such as in the .rodata section. This data blob is the true payload of the malware, but it has been put through a process such as encryption, compression, or another form of manipulation to render it unreadable. Sometimes, this data blob is instead shipped in a different file, such as a .dat file, or a fake image. When this happens, the main loader has to read this using a syscall, which can be useful for analysis as syscalls can be easily identified, even in heavily obfuscated binaries.
In the main executable, malware developers will typically include an unpacking stub that takes the data blob, performs one or more operations on it, and then triggers its execution. In most samples, the decoded payload data is loaded into a newly allocated memory region, which will then be marked as executable and executed. In other cases, the decoded data is instead dropped into a new executable on disk and run, but this is less common as it increases the likelihood of detection.
Finding the unpacking routine
The first stage of analysis is uncovering the unpacking routine so it can be reverse engineered. There are several ways to approach this, but it is traditionally first triaged via static analysis on the initial stages available to the analyst.
SnappyBee consists of two components that can be analyzed:
A Dynamic-link Library (DLL) that acts as a loader, responsible for unpacking the malicious code
A data file shipped alongside the DLL, which contains the encrypted malicious code
Additionally, SnappyBee includes a legitimate signed executable that is vulnerable to DLL side-loading. This means that when the executable is run, it will inadvertently load SnappyBee’s DLL instead of the legitimate one it expects. This allows SnappyBee to appear more legitimate to antivirus solutions.
The first stage of analysis is performing static analysis of the DLL. This can be done by opening the DLL within a disassembler such as IDA Pro. Upon opening the DLL, IDA will display the DllMain function, which is the malware’s initial entry point and the first code executed when the DLL is loaded.
Figure 1: The DllMain function
First, the function checks if the variable fdwReason is set to 1, and exits if it is not. This variable is set by Windows to indicate why the DLL was loaded. According to Microsoft Developer Network (MSDN), a value of 1 corresponds to DLL_PROCESS_ATTACH, meaning “The DLL is being loaded into the virtual address space of the current process as a result of the process starting up or as a result of a call to LoadLibrary” [3]. Since SnappyBee is known to use DLL sideloading for execution, DLL_PROCESS_ATTACH is the expected value when the legitimate executable loads the malicious DLL.
SnappyBee then uses the GetModule and GetProcAddress to dynamically resolve the address of the VirtualProtect in kernel32 and StartServiceCtrlDispatcherW in advapi32. Resolving these dynamically at runtime prevents them from showing up as a static import for the module, which can help evade detection by anti-malware solutions. Different regions of memory have different permissions to control what they can be used for, with the main ones being read, write, and execute. VirtualProtect is a function that changes the permissions of a given memory region.
SnappyBee then uses VirtualProtect to set the memory region containing the code for the StartServiceCtrlDispatcherW function as writable. It then inserts a jump instruction at the start of this function, redirecting the control flow to one of the SnappyBee DLL’s other functions, and then restores the old permissions.
In practice, this means when the legitimate executable calls StartServiceCtrlDispatcherW, it will immediately hand execution back to SnappyBee. Meanwhile, the call stack now appears more legitimate to outside observers such as antimalware solutions.
The hooked-in function then reads the data file that is shipped with SnappyBee and loads it into a new memory allocation. This pattern of loading the file into memory likely means it is responsible for unpacking the next stage.
Figure 2: The start of the unpacking routine that reads in dbindex.dat.
SnappyBee then proceeds to decrypt the memory allocation and execute the code.
Figure 3: The memory decryption routine.
This section may look complex, however it is fairly straight forward. Firstly, it uses memset to zero out a stack variable, which will be used to store the decryption key. It then uses the first 16 bytes of the data file as a decryption key to initialize the context from.
SnappyBee then calls the mbed_tls_arc4_crypt function, which is a function from the mbedtls library. Documentation for this function can be found online and can be referenced to better understand what each of the arguments mean [4].
Figure 4: The documentation for mbedtls_arc4_ crypt.
Comparing the decompilation with the documentation, the arguments SnappyBee passes to the function can be decoded as:
The context derived from 16-byte key at the start of the data is passed in as the context in the first parameter
The file size minus 16 bytes (to account for the key at the start of the file) is the length of the data to be decrypted
A pointer to the file contents in memory, plus 16 bytes to skip the key, is used as the input
A pointer to a new memory allocation obtained from VirtualAlloc is used as the output
So, putting it all together, it can be concluded that SnappyBee uses the first 16 bytes as the key to decrypt the data that follows , writing the output into the allocated memory region.
SnappyBee then calls VirtualProtect to set the decrypted memory region as Read+Execute, and subsequently executes the code at the memory pointer. This is clearly where the unpacked code containing the next stage will be placed.
Unpacking the malware
Understanding how the unpacking routine works is the first step. The next step is obtaining the actual code, which cannot be achieved through static analysis alone.
There are two viable methods to retrieve the next stage. The first method is implementing the unpacking routine from scratch in a language like Python and running it against the data file.
This is straightforward in this case, as the unpacking routine in relatively simple and would not require much effort to re-implement. However, many unpacking routines are far more complex, which leads to the second method: allowing the malware to unpack itself by debugging it and then capturing the result. This is the approach many analysts take to unpacking, and the following will document this method to unpack SnappyBee.
As SnappyBee is 32-bit Windows malware, debugging can be performed using x86dbg in a Windows sandbox environment to debug SnappyBee. It is essential this sandbox is configured correctly, because any mistake during debugging could result in executing malicious code, which could have serious consequences.
Before debugging, it is necessary to disable the DYNAMIC_BASE flag on the DLL using a tool such as setdllcharacteristics. This will stop ASLR from randomizing the memory addresses each time the malware runs and ensures that it matches the addresses observed during static analysis.
The first place to set a breakpoint is DllMain, as this is the start of the malicious code and the logical place to pause before proceeding. Using IDA, the functions address can be determined; in this case, it is at offset 10002DB0. This can be used in the Goto (CTRL+G) dialog to jump to the offset and place a breakpoint. Note that the “Run to user code” button may need to be pressed if the DLL has not yet been loaded by x32dbg, as it spawns a small process to load the DLL as DLLs cannot be executed directly.
The program can then run until the breakpoint, at which point the program will pause and code recognizable from static analysis can be observed.
Figure 5: The x32dbg dissassembly listing forDllMain.
In the previous section, this function was noted as responsible for setting up a hook, and in the disassembly listing the hook address can be seen being loaded at offset 10002E1C. It is not necessary to go through the whole hooking process, because only the function that gets hooked in needs to be run. This function will not be naturally invoked as the DLL is being loaded directly rather than via sideloading as it expects. To work around this, the Extended Instruction Pointer (EIP) register can be manipulated to point to the start of the hook function instead, which will cause it to run instead of the DllMain function.
To update EIP, the CRTL+G dialog can again be used to jump to the hook function address (10002B50), and then the EIP register can be set to this address by right clicking the first instruction and selecting “Set EIP here”. This will make the hook function code run next.
Figure 6: The start of the hookedin-in function
Once in this function, there are a few addresses where breakpoints should be set in order to inspect the state of the program at critical points in the unpacking process. These are:
- 10002C93, which allocates the memory for the data file and final code
- 10002D2D, which decrypts the memory
- 10002D81, which runs the unpacked code
Setting these can be done by pressing the dot next to the instruction listing, or via the CTRL+G Goto menu.
At the first breakpoint, the call to VirtualAlloc will be executed. The function returns the memory address of the created memory region, which is stored in the EAX register. In this case, the region was allocated at address 00700000.
Figure 7: The result of the VirtualAlloc call.
It is possible to right click the address and press “Follow in dump” to pin the contents of the memory to the lower pane, which makes it easy to monitor the region as the unpacking process continues.
Figure 8: The allocated memory region shown in x32dbg’s dump.
Single-stepping through the application from this point eventually reaches the call to ReadFile, which loads the file into the memory region.
Figure 9: The allocated memory region after the file is read into it, showing high entropy data.
The program can then be allowed to run until the next breakpoint, which after single-stepping will execute the call to mbedtls_arc4_crypt to decrypt the memory. At this point, the data in the dump will have changed.
Figure 10: The same memory region after the decryption is run, showing lower entropy data.
Right-clicking in the dump and selecting "Disassembly” will disassemble the data. This yields valid shell code, indicating that the unpacking succeeded, whereas corrupt or random data would be expected if the unpacking had failed.
Figure 11: The disassembly view of the allocated memory.
Right-clicking and selecting “Follow in memory map” will show the memory allocation under the memory map view. Right-clicking this then provides an option to dump the entire memory block to file.
Figure 12: Saving the allocated memory region.
This dump can then be opened in IDA, enabling further static analysis of the shellcode. Reviewing the shellcode, it becomes clear that it performs another layer of unpacking.
As the debugger is already running, the sample can be allowed to execute up to the final breakpoint that was set on the call to the unpacked shellcode. Stepping into this call will then allow debugging of the new shellcode.
The simplest way to proceed is to single-step through the code, pausing on each call instruction to consider its purpose. Eventually, a call instruction that points to one of the memory regions that were assigned will be reached, which will contain the next layer of unpacked code. Using the same disassembly technique as before, it can be confirmed that this is more unpacked shellcode.
Figure 13: The unpacked shellcode’s call to RDI, which points to more unpacked shellcode. Note this screenshot depicts the 64-bit variant of SnappyBee instead of 32-bit, however the theory is the same.
Once again, this can be dumped out and analyzed further in IDA. In this case, it is the final payload used by the SnappyBee malware.
Conclusion
Unpacking remains one of the most common anti-analysis techniques and is a feature of most sophisticated malware from threat groups. This technique of in-memory decryption reduces the forensic “surface area” of the malware, helping it to evade detection from anti-malware solutions. This blog walks through one such example and provides practical knowledge on how to unpack malware for deeper analysis.
In addition, this blog has detailed several other techniques used by threat actors to evade analysis, such as DLL sideloading to execute code without arising suspicion, dynamic API resolving to bypass static heuristics, and multiple nested stages to make analysis challenging.
Malware such as SnappyBee demonstrates a continued shift towards highly modular and low-friction malware toolkits that can be reused across many intrusions and campaigns. It remains vital for security teams to maintain the ability to combat the techniques seen in these toolkits when responding to infections.
While the technical details of these techniques are primarily important to analysts, the outcomes of this work directly affect how a Security Operations Centre (SOC) operates at scale. Without the technical capability to reliably unpack and observe these samples, organizations are forced to respond without the full picture.
The techniques demonstrated here help close that gap. This enables security teams to reduce dwell time by understanding the exact mechanisms of a sample earlier, improve detection quality with behavior-based indicators rather than relying on hash-based detections, and increase confidence in response decisions when determining impact.
Credit to Nathaniel Bill (Malware Research Engineer) Edited by Ryan Traill (Analyst Content Lead)
The State of AI Cybersecurity 2026: Unveiling insights from over 1,500 security leaders
2025 was the year enterprise AI went mainstream. In 2026, it’s made its way into every facet of the organizational structure – transforming workflows, revolutionizing productivity, and creating new value streams. In short, it’s opened up a whole new attack surface.
At the same time, AI has accelerated the pace of cybersecurity arms race on both sides: adversaries are innovating using the latest AI technologies at their disposal while defenders scramble to outmaneuver them and stay ahead of AI-powered threats.
That’s why Darktrace publishes this research every year. The State of AI Cybersecurity 2026 provides an annual snapshot of how the AI threat landscape is shifting, where organizations are adopting AI to maximum advantage, and how they are securing AI in the enterprise.
What is the State of AI Cybersecurity 2026?
We surveyed over 1,500 CISOs, IT leaders, administrators, and practitioners from a range of industries and different countries to uncover their attitudes, understanding, and priorities when it comes to AI threats, agents, tools, and operations in 2026.
The results show a fast-changing picture, as security leaders race to navigate the challenges and opportunities at play. Since last year, there has been enormous progress towards maturity in areas like AI literacy and confidence in AI-powered defense, while issues around AI governance remain inconclusive.
Let’s look at some of the key findings for 2026.
What’s the impact of AI on the attack surface?
Security leaders are seeing the adoption of AI agents across the workforce, and are increasingly concerned about the security implications.
44% are extremely or very concerned with the security implications of third-party LLMs (like Copilot or ChatGPT)
92% are concerned about the use of AI agents across the workforce and their impact on security
The rapid expansion of generative AI across the enterprise is outpacing the security frameworks designed to govern it. AI systems behave in ways that traditional defenses are not designed to monitor, introducing new risks around data exposure, unauthorized actions, and opaque decision-making as employees embed generative AI and autonomous agents into everyday workflows.
Their top concerns? Sensitive data exposure ranks top (61%), while regulatory compliance violations are a close second (56%). These risks tend to have the fastest and most material fallout – ranging from fines to reputational harm – and are more likely to materialize in environments where AI governance is still evolving.
What’s the impact of AI on the cyber threat landscape?
AI is now being used to expedite every stage of the attack kill chain – from initial intrusion to privilege escalation and data exfiltration.
“73% say that AI-powered threats are already having a significant impact on their organization.”
With AI, attackers can launch novel attacks at scale, and this is significantly increasing the number of threats requiring attention by the security team – often to the point of overwhelm.
Traditional security solutions relying on historical attack data were never designed to handle an environment where attacks continuously evolve, multiply, and optimize at machine speed, so it’s no surprise that 92% agree that AI-powered cyber-threats are forcing them to significantly upgrade their defenses.
How is AI reshaping cybersecurity operations?
Cybersecurity workflows are still in flux as security leaders get used to the integration of AI agents into everyday operations.
“Generative AI is now playing a role in 77% of security stacks.” But only 35% are using unsupervised machine learning.
AI technologies are diverse, ranging from LLMs to NLP systems, GANs, and unsupervised machine learning, with each type offering specific capabilities and facing particular limitations. The lack of familiarity with the different types of AI used within the security stack may be holding some practitioners back from using these new technologies to their best advantage.
It also creates a lack of trust between humans and AI systems: only 14% of security professionals allow AI to take independent remediation actions in the SOC with no human in the loop.
Another new trend for this year is a strong preference (85%) for relying on Managed Security Service Providers (MSSPs) for SOC services instead of in-house teams, as organizations aim to secure expert, always-on support without the cost and operational burden of running an internal operation.
What impact is AI having on cybersecurity tools?
“96% of cybersecurity professionals agree that AI can significantly improve the speed and efficiency with which they work.”
The capacity of AI for augmenting security efforts is undisputed. But as vendor AI claims become far-reaching, it falls to security leaders to clarify which AI tools offer true value and can help solve their specific security challenges.
Security professionals are aligned on the biggest area of impact: 72% agree that AI excels at detecting anomalies thanks to its advanced pattern recognition. This enables it to identify unusual behavior that may signal a threat, even when the specific attack has never been encountered or recorded in existing datasets.
“When purchasing new security capabilities, 93% prefer ones that are part of a broader platform over individual point products.”
Like last year, the drive towards platform consolidation remains strong. Fewer vendors can mean tighter integrations, less console switching, streamlined management, and stronger cross-domain threat insights. The challenge is finding vendors that perform well across the board.
See the full report for more statistics and insights into how security leaders are responding to the AI landscape in 2026.
.jpg)
