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.
React2Shell: How Opportunist Attackers Exploited CVE-2025-55182 Within Hours
What is React2Shell?
CVE-2025-55182, also known as React2Shell is a vulnerability within React server components that allows for an unauthenticated attacker to gain remote code execution with a single request. The severity of this vulnerability and ease of exploitability has led to threat actors opportunistically exploiting it within a matter of days of its public disclosure.
Darktrace security researchers rapidly deployed a new honeypot using the Cloudypots system, allowing for the monitoring of exploitation of the vulnerability in the wild.
Cloudypots is a system that enables virtual instances of vulnerable applications to be deployed in the cloud and monitored for attack. This approach allows for Darktrace to deploy high-interaction, realistic honeypots, that appear as genuine deployments of vulnerable software to attackers.
This blog will explore one such campaign, nicknamed “Nuts & Bolts” based on the naming used in payloads.
Analysis of the React2Shell exploit
The React2Shell exploit relies on an insecure deserialization vulnerability within React Server Components’ “Flight” protocol. This protocol uses a custom serialization scheme that security researchers discovered could be abused to run arbitrary JavaScript by crafting the serialized data in a specific way. This is possible because the framework did not perform proper type checking, allowing an attacker to reference types that can be abused to craft a chain that resolves to an anonymous function, and then invoke it with the desired JavaScript as a promise chain.
This code execution can then be used to load the ‘child_process’ node module and execute any command on the target server.
The vulnerability was discovered on December 3, 2025 with a patch made available on December 3 [1]. Within 30 hours of the patch, a publicly available proof of concept emerged that could be used to exploit any vulnerable server. This rapid timeline left many servers remaining unpatched by the time attackers began actively exploiting the vulnerability.
Initial access
The threat actor behind the “Nuts & Bolts” campaign uses a spreader server with IP 95.214.52[.]170 to infect victims. The IP appears to be located in Poland and is associated with a hosting provided known as MEVSPACE. The spreader is highly aggressive, launching exploitation attempts, roughly every hour.
When scanning, he spreader primarily targets port 3000, which is the default port for a NEXT.js server in a default or development configuration. It is possible the attacker is avoiding port 80 and 443, as these are more likely to have reverse proxies or WAFs in front of the server, which could disrupt exploitation attempts.
When the spreader finds a new host with port 3000 open, it begins by testing if it is vulnerable to React2Shell by sending a crafted request to run the ‘whoami’ command and store the output in an error digest that is returned to the attacker.
The above snippet is the core part of the crafted request that performs the execution. This allows the attacker to confirm that the server is vulnerable and fetch the user account under which the NEXT.js process is running, which is useful information for determining if a target is worth attacking.
From here, the attacker then sends an additional request to run the actual payload on the victim server.
This snippet attempts to deploy several payloads by using wget (or curl if wget fails) into the /dev directory and execute them. The x86 binary is a Mirai variant that does not appear to have any major alterations to regular Mirai. The ‘nuts/bolts’ endpoint returns a bash script, which is then executed. The script includes several log statements throughout its execution to provide visibility into which parts ran successfully. Similar to the ‘whoami’ request, the output is placed in an error digest for the attacker to review.
In this case, the command-and-control (C2) IP, 89[.]144.31.18, is hosted on a different server operated by a German hosting provider named myPrepaidServer, which offers virtual private server (VPS) services and accepts cryptocurrency payments [2].
Figure 1: Logs observed in the NEXT.JS console as a result of exploitation. In this case, the honeypot was attacked just two minutes after being deployed.
Nuts & Bolts script
This script’s primary purpose is to prepare the box for a cryptocurrency miner.
The script starts by attempting to terminate any competing cryptocurrency miner processes using ‘pkill’ that match on a specific name. It will check for and terminate:
xmrig
softirq (this also matches a system process, which it will fail to kill each invocation)
watcher
/tmp/a.sh
health.sh
Following this, the script will checks for a process named “fghgf”. If it is not running, it will retrieve hxxp://89[.]144.31.18/nuts/lc and write it to /dev/ijnegrrinje.json, as well as retrieving hxxp://89[.]144.31.18/nuts/x and writing it to /dev/fghgf. The script will the executes /dev/fghgf -c /dev/ijnegrrinje.json -B in the background, which is an XMRig miner.
Figure 2: The XMRig deployment script.
The miner is configured to connect to two private pools at 37[.]114.37.94 and 37[.]114.37.82, using “poop” as both the username and password. The use of a private pool conceals the associated wallet address. From here, a short bash script is dropped to /dev/stink.sh. This script continuously crawls all running processes on the system and reads their /proc/pid/exe path, which contains a copy of the original executable that was run. The ‘strings’ utility is run to output all valid ASCII strings found within the data and checks to see if contains either “xmrig”, “rondo” or “UPX 5”. If so, it sends a SIGKILL to the process to terminate it.
Additionally, it will run ‘ls –l’ on the exe path in case it is symlinked to a specific path or has been deleted. If the output contains any of the following strings, the script sends a SIGKILL to terminate the program:
(deleted) - Indicates that the original executable was deleted from the disk, a common tactic used by malware to evade detection.
xmrig
hash
watcher
/dev/a
softirq
rondo
UPX 5.02
Figure 3: The killer loop and the dropper. In this case ${R}/${K} resolves to /dev/stink.sh.
Darktrace observations in customer environments
Following the public disclosure of CVE‑2025‑55182 on December 3rd, 2025 Darktrace observed multiple exploitation attempts across customer environments beginning around December 4. Darktrace triage identified a series of consistent indicators of compromise (IoCs). By consolidating indicators across multiple deployments and repeat infrastructure clusters, Darktrace identified a consistent kill chain involving shell‑script downloads and HTTP beaconing.
In one example, on December 5, Darktrace observed external connections to malicious IoC endpoints (172.245.5[.]61:38085, 5.255.121[.]141, 193.34.213[.]15), followed by additional connections to other potentially malicious endpoint. These appeared related to the IoCs detailed above, as one suspicious IP address shared the same ASN. After this suspicious external connectivity, Darktrace observed cryptomining-related activity. A few hours later, the device initiated potential lateral movement activity, attempting SMB and RDP sessions with other internal devices on the network. These chain of events appear to identify this activity to be related to the malicious campaign of the exploitation of React2Shell vulnerability.
Generally, outbound HTTP traffic was observed to ports in the range of 3000–3011, most notably port 3001. Requests frequently originated from scripted tools, with user agents such as curl/7.76.1, curl/8.5.0, Wget/1.21.4, and other generic HTTP signatures. The URIs associated with these requests included paths like /nuts/x86 and /n2/x86, as well as long, randomized shell script names such as /gfdsgsdfhfsd_ghsfdgsfdgsdfg.sh. In some cases, parameterized loaders were observed, using query strings like: /?h=<ip>&p=<port>&t=<proto>&a=l64&stage=true.
Infrastructure analysis revealed repeated callbacks to IP-only hosts linked to ASN AS200593 (Prospero OOO), a well-known “bulletproof” hosting provider often utilized by cyber criminals [3], including addresses such as 193.24.123[.]68:3001 and 91.215.85[.]42:3000, alongside other nodes hosting payloads and staging content.
Darktrace model coverage
Darktrace model coverage consistently highlighted behaviors indicative of exploitation. Among the most frequent detections were anomalous server activity on new, non-standard ports and HTTP requests posted to IP addresses without hostnames, often using uncommon application protocols. Models also flagged the appearance of new user agents such as curl and wget originating from internet-facing systems, representing an unusual deviation from baseline behavior.
Additionally, observed activity included the download of scripts and executable files from rare external sources, with Darktrace’s Autonomous Response capability intervening to block suspicious transfers, when enabled. Beaconing patterns were another strong signal, with detections for HTTP beaconing to new or rare IP addresses, sustained SSL or HTTP increases, and long-running compromise indicators such as “Beacon for 4 Days” and “Slow Beaconing.”
Conclusion
While this opportunistic campaign to exploit the React2Shell exploit is not particularly sophisticated, it demonstrates that attackers can rapidly prototyping new methods to take advantage of novel vulnerabilities before widespread patching occurs. With a time to infection of only two minutes from the initial deployment of the honeypot, this serves as a clear reminder that patching vulnerabilities as soon as they are released is paramount.
Credit to Nathaniel Bill (Malware Research Engineer), George Kim (Analyst Consulting Lead – AMS), Calum Hall (Technical Content Researcher), Tara Gould (Malware Research Lead, and Signe Zaharka (Principal Cyber Analyst).
Organizations are built on increasingly complex digital estates. Nowadays, the average IT ecosystem spans across a large web of interconnected domains like identity, network, cloud, and email.
While these domain-specific technologies may boost business efficiency and scalability, they also provide blind spots where attackers can shelter undetected. Threat actors can slip past defenses because security teams often use different detection tools in each realm of their digital infrastructure. Adversaries will purposefully execute different stages of an attack across different domains, ensuring no single tool picks up too many traces of their malicious activity. Identifying and investigating this type of threat, known as a cross-domain attack, requires mastery in event correlation.
For example, one isolated network scan detected on your network may seem harmless at first glance. Only when it is stitched together with a rare O365 login, a new email rule and anomalous remote connections to an S3 bucket in AWS does it begin to manifest as an actual intrusion.
However, there are a whole host of other challenges that arise with detecting this type of attack. Accessing those alerts in the respective on-premise network, SaaS and IaaS environments, understanding them and identifying which ones are related to each other takes significant experience, skill and time. And time favours no one but the threat actor.
Figure 1: Anatomy of a cross domain attack
Diverse domains and empty grocery shelves
In April 2025, the UK faced a throwback to pandemic-era shortages when the supermarket giant Marks & Spencer (M&S) was crippled by a cyberattack, leaving empty shelves across its stores and massive disruptions to its online service.
The threat actors, a group called Scattered Spider, exploited multiple layers of the organization’s digital infrastructure. Notably, the group were able to bypass the perimeter not by exploiting a technical vulnerability, but an identity. They used social engineering tactics to impersonate an M&S employee and successfully request a password reset.
Once authenticated on the network, they accessed the Windows domain controller and exfiltrated the NTDS.dit file – a critical file containing hashed passwords for all users in the domain. After cracking those hashes offline, they returned to the network with escalated privileges and set their sights on the M&S cloud infrastructure. They then launched the encryption payload on the company’s ESXi virtual machines.
To wrap up, the threat actors used a compromised employee’s email account to send an “abuse-filled” email to the M&S CEO, bragging about the hack and demanding payment. This was possibly more of a psychological attack on the CEO than a technically integral part of the cyber kill chain. However, it revealed yet another one of M&S’s domains had been compromised.
In summary, the group’s attack spanned four different domains:
Identity: Social engineering user impersonation
Network: Exfiltration of NTDS.dit file
Cloud: Ransomware deployed on ESXI VMs
Email: Compromise of user account to contact the CEO
Adept at exploiting nuance
This year alone, several high-profile cyber-attacks have been attributed to the same group, Scattered Spider, including the hacks on Victoria’s Secret, Adidas, Hawaiian Airlines, WestJet, the Co-op and Harrods. It begs the question, what has made this group so successful?
In the M&S attack, they showcased their advanced proficiency in social engineering, which they use to bypass identity controls and gain initial access. They demonstrated deep knowledge of cloud environments by deploying ransomware onto virtualised infrastructure. However, this does not exemplify a cookie-cutter template of attack methods that brings them success every time.
According to CISA, Scattered Spider typically use a remarkable variety of TTPs (tactics, techniques and procedures) across multiple domains to carry out their campaigns. From leveraging legitimate remote access tools in the network, to manipulating AWS EC2 cloud instances or spoofing email domains, the list of TTPs used by the group is eye-wateringly long. Additionally, the group reportedly evades detection by “frequently modifying their TTPs”.
If only they had better intentions. Any security director would be proud of a red team who not only has this depth and breadth of domain-centric knowledge but is also consistently upskilling.
Yet, staying ahead of adversaries who seamlessly move across domains and fluently exploit every system they encounter is just one of many hurdles security teams face when investigating cross-domain attacks.
Resource-heavy investigations
There was a significant delay in time to detection of the M&S intrusion. News outlet BleepingComputer reported that attackers infiltrated the M&S network as early as February 2025. They maintained persistence for weeks before launching the attack in late April 2025, indicating that early signs of compromise were missed or not correlated across domains.
While it’s unclear exactly why M&S missed the initial intrusion, one can speculate about the unique challenges investigating cross-domain attacks present.
Challenges of cross-domain investigation
First and foremost, correlation work is arduous because the string of malicious behaviour doesn’t always stem from the same device.
A hypothetical attack could begin with an O365 credential creating a new email rule. Weeks later, that same credential authenticates anomalously on two different devices. One device downloads an .exe file from a strange website, while the other starts beaconing every minute to a rare external IP address that no one else in the organisation has ever connected to. A month later, a third device downloads 1.3 GiB of data from a recently spun up S3 bucket and gradually transfers a similar amount of data to that same rare IP.
Amid a sea of alerts and false positives, connecting the dots of a malicious attack like this takes time and meticulous correlation. Factor in the nuanced telemetry data related to each domain and things get even more complex.
An analyst who specialises in network security may not understand the unique logging formats or API calls in the cloud environment. Perhaps they are proficient in protecting the Windows Active Directory but are unfamiliar with cloud IAM.
Cloud is also an inherently more difficult domain to investigate. With 89% of organizations now operating in multi-cloud environments time must be spent collecting logs, snapshots and access records. Coupled with the threat of an ephemeral asset disappearing, the risk of missing a threat is high. These are some of the reasons why research shows that 65% of organisations spend 3-5 extra days investigating cloud incidents.
Helpdesk teams handling user requests over the phone require a different set of skills altogether. Imagine a threat actor posing as an employee and articulately requesting an urgent password reset or a temporary MFA deactivation. The junior Helpdesk agent— unfamiliar with the exception criteria, eager to help and feeling pressure from the persuasive manipulator at the end of the phoneline—could easily fall victim to this type of social engineering.
Empowering analysts through intelligent automation
Even the most skilled analysts can’t manually piece together every strand of malicious activity stretching across domains. But skill alone isn’t enough. The biggest hurdle in investigating these attacks often comes down to whether the team have the time, context, and connected visibility needed to see the full picture.
Many organizations attempt to bridge the gap by stitching together a patchwork of security tools. One platform for email, another for endpoint, another for cloud, and so on. But this fragmentation reinforces the very silos that cross-domain attacks exploit. Logs must be exported, normalized, and parsed across tools a process that is not only error-prone but slow. By the time indicators are correlated, the intrusion has often already deepened.
That’s why automation and AI are becoming indispensable. The future of cross-domain investigation lies in systems that can:
Automatically correlate activity across domains and data sources, turning disjointed alerts into a single, interpretable incident.
Generate and test hypotheses autonomously, identifying likely chains of malicious behaviour without waiting for human triage.
Explain findings in human terms, reducing the knowledge gap between junior and senior analysts.
Operate within and across hybrid environments, from on-premise networks to SaaS, IaaS, and identity systems.
This is where Darktrace transforms alerting and investigations. Darktrace’s Cyber AI Analyst automates the process of correlation, hypothesis testing, and narrative building, not just within one domain, but across many. An anomalous O365 login, a new S3 bucket, and a suspicious beaconing host are stitched together automatically, surfacing the story behind the alerts rather than leaving it buried in telemetry.
Figure 2: How threat activity is correlated in Cyber AI Analyst
By analyzing events from disparate tools and sources, AI Analyst constructs a unified timeline of activity showing what happened, how it spread, and where to focus next. For analysts, it means investigation time is measured in minutes, not days. For security leaders, it means every member of the SOC, regardless of experience, can contribute meaningfully to a cross-domain response.
Figure 3: Correlation showcasing cross domains (SaaS and IaaS) in Cyber AI Analyst
Until now, forensic investigations were slow, manual, and reserved for only the largest organizations with specialized DFIR expertise. Darktrace / Forensic Acquisition & Investigation changes that by leveraging the scale and elasticity of the cloud itself to automate the entire investigation process. From capturing full disk and memory at detection to reconstructing attacker timelines in minutes, the solution turns fragmented workflows into streamlined investigations available to every team.
What once took days now takes minutes. Now, forensic investigations in the cloud are faster, more scalable, and finally accessible to every security team, no matter their size or expertise.
