Blog
/
AI
/
April 16, 2025

Introducing Version 2 of Darktrace’s Embedding Model for Investigation of Security Threats (DEMIST-2)

Learn how Darktrace’s DEMIST-2 embedding model delivers high-accuracy threat classification and detection across any environment, outperforming larger models with efficiency and precision.
Inside the SOC
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
Margaret Cunningham, PhD
Director, Security & AI Strategy, Field CISO
woman looking at laptop at deskDefault blog imageDefault blog imageDefault blog imageDefault blog imageDefault blog imageDefault blog image
16
Apr 2025

DEMIST-2 is Darktrace’s latest embedding model, built to interpret and classify security data with precision. It performs highly specialized tasks and can be deployed in any environment. Unlike generative language models, DEMIST-2 focuses on providing reliable, high-accuracy detections for critical security use cases.

DEMIST-2 Core Capabilities:  

  • Enhances Cyber AI Analyst’s ability to triage and reason about security incidents by providing expert representation and classification of security data, and as a part of our broader multi-layered AI system
  • Classifies and interprets security data, in contrast to language models that generate unpredictable open-ended text responses  
  • Incorporates new innovations in language model development and architecture, optimized specifically for cybersecurity applications
  • Deployable across cloud, on-prem, and edge environments, DEMIST-2 delivers low-latency, high-accuracy results wherever it runs. It enables inference anywhere.

Cybersecurity is constantly evolving, but the need to build precise and reliable detections remains constant in the face of new and emerging threats. Darktrace’s Embedding Model for Investigation of Security Threats (DEMIST-2) addresses these critical needs and is designed to create stable, high-fidelity representations of security data while also serving as a powerful classifier. For security teams, this means faster, more accurate threat detection with reduced manual investigation. DEMIST-2's efficiency also reduces the need to invest in massive computational resources, enabling effective protection at scale without added complexity.  

As an embedding language model, DEMIST-2 classifies and creates meaning out of complex security data. This equips our Self-Learning AI with the insights to compare, correlate, and reason with consistency and precision. Classifications and embeddings power core capabilities across our products where accuracy is not optional, as a part of our multi-layered approach to AI architecture.

Perhaps most importantly, DEMIST-2 features a compact architecture that delivers analyst-level insights while meeting diverse deployment needs across cloud, on-prem, and edge environments. Trained on a mixture of general and domain-specific data and designed to support task specialization, DEMIST-2 provides privacy-preserving inference anywhere, while outperforming larger general-purpose models in key cybersecurity tasks.

This proprietary language model reflects Darktrace's ongoing commitment to continually innovate our AI solutions to meet the unique challenges of the security industry. We approach AI differently, integrating diverse insights to solve complex cybersecurity problems. DEMIST-2 shows that a refined, optimized, domain-specific language model can deliver outsized results in an efficient package. We are redefining possibilities for cybersecurity, but our methods transfer readily to other domains. We are eager to share our findings to accelerate innovation in the field.  

The evolution of DEMIST-2

Key concepts:  

  • Tokens: The smallest units processed by language models. Text is split into fragments based on frequency patterns allowing models to handle unfamiliar words efficiently
  • Low-Rank Adaptors (LoRA): Small, trainable components added to a model that allow it to specialize in new tasks without retraining the full system. These components learn task-specific behavior while the original foundation model remains unchanged. This approach enables multiple specializations to coexist, and work simultaneously, without drastically increasing processing and memory requirements.

Darktrace began using large language models in our products in 2022. DEMIST-2 reflects significant advancements in our continuous experimentation and adoption of innovations in the field to address the unique needs of the security industry.  

It is important to note that Darktrace uses a range of language models throughout its products, but each one is chosen for the task at hand. Many others in the artificial intelligence (AI) industry are focused on broad application of large language models (LLMs) for open-ended text generation tasks. Our research shows that using LLMs for classification and embedding offers better, more reliable, results for core security use cases. We’ve found that using LLMs for open-ended outputs can introduce uncertainty through inaccurate and unreliable responses, which is detrimental for environments where precision matters. Generative AI should not be applied to use cases, such as investigation and threat detection, where the results can deeply matter. Thoughtful application of generative AI capabilities, such as drafting decoy phishing emails or crafting non-consequential summaries are helpful but still require careful oversight.

Data is perhaps the most important factor for building language models. The data used to train DEMIST-2 balanced the need for general language understanding with security expertise. We used both publicly available and proprietary datasets.  Our proprietary dataset included privacy-preserving data such as URIs observed in customer alerts, anonymized at source to remove PII and gathered via the Call Home and aianalyst.darktrace.com services. For additional details, read our Technical Paper.  

DEMIST-2 is our way of addressing the unique challenges posed by security data. It recognizes that security data follows its own patterns that are distinct from natural language. For example, hostnames, HTTP headers, and certificate fields often appear in predictable ways, but not necessarily in a way that mirrors natural language. General-purpose LLMs tend to break down when used in these types of highly specialized domains. They struggle to interpret structure and context, fragmenting important patterns during tokenization in ways that can have a negative impact on performance.  

DEMIST-2 was built to understand the language and structure of security data using a custom tokenizer built around a security-specific vocabulary of over 16,000 words. This tokenizer allows the model to process inputs more accurately like encoded payloads, file paths, subdomain chains, and command-line arguments. These types of data are often misinterpreted by general-purpose models.  

When the tokenizer encounters unfamiliar or irregular input, it breaks the data into smaller pieces so it can still be processed. The ability to fall back to individual bytes is critical in cybersecurity contexts where novel or obfuscated content is common. This approach combines precision with flexibility, supporting specialized understanding with resilience in the face of unpredictable data.  

Along with our custom tokenizer, we made changes to support task specialization without increasing model size. To do this, DEMIST-2 uses LoRA . LoRA is a technique that integrates lightweight components with the base model to allow it to perform specific tasks while keeping memory requirements low. By using LoRA, our proprietary representation of security knowledge can be shared and reused as a starting point for more highly specialized models, for example, it takes a different type of specialization to understand hostnames versus to understand sensitive filenames. DEMIST-2 dynamically adapts to these needs and performs them with purpose.  

The result is that DEMIST-2 is like having a room of specialists working on difficult problems together, while sharing a basic core set of knowledge that does not need to be repeated or reintroduced to every situation. Sharing a consistent base model also improves its maintainability and allows efficient deployment across diverse environments without compromising speed or accuracy.  

Tokenization and task specialization represent only a portion of the updates we have made to our embedding model. In conjunction with the changes described above, DEMIST-2 integrates several updated modeling techniques that reduce latency and improve detections. To learn more about these details, our training data and methods, and a full write-up of our results, please read our scientific whitepaper.

DEMIST-2 in action

In this section, we highlight DEMIST-2's embeddings and performance. First, we show a visualization of how DEMIST-2 classifies and interprets hostnames, and second, we present its performance in a hostname classification task in comparison to other language models.  

Embeddings can often feel abstract, so let’s make them real. Figure 1 below is a 2D visualization of how DEMIST-2 classifies and understands hostnames. In reality, these hostnames exist across many more dimensions, capturing details like their relationships with other hostnames, usage patterns, and contextual data. The colors and positions in the diagram represent a simplified view of how DEMIST-2 organizes and interprets these hostnames, providing insights into their meaning and connections. Just like an experienced human analyst can quickly identify and group hostnames based on patterns and context, DEMIST-2 does the same at scale.  

DEMIST-2 visualization of hostname relationships from a large web dataset.
Figure 1: DEMIST-2 visualization of hostname relationships from a large web dataset.

Next, let’s zoom in on two distinct clusters that DEMIST-2 recognizes. One cluster represents small businesses (Figure 2) and the other, Russian and Polish sites with similar numerical formats (Figure 3). These clusters demonstrate how DEMIST-2 can identify specific groupings based on real-world attributes such as regional patterns in website structures, common formats used by small businesses, and other properties such as its understanding of how websites relate to each other on the internet.

Cluster of small businesses
Figure 2: Cluster of small businesses
Figure 3: Cluster of Russian and Polish sites with a similar numerical format

The previous figures provided a view of how DEMIST-2 works. Figure 4 highlights DEMIST-2’s performance in a security-related classification task. The chart shows how DEMIST-2, with just 95 million parameters, achieves nearly 94% accuracy—making it the highest-performing model in the chart, despite being the smallest. In comparison, the larger model with 278 million parameters achieves only about 89% accuracy, showing that size doesn’t always mean better performance. Small models don’t mean poor performance. For many security-related tasks, DEMIST-2 outperforms much larger models.

Hostname classification task performance comparison against comparable open source foundation models
Figure 4: Hostname classification task performance comparison against comparable open source foundation models

With these examples of DEMIST-2 in action, we’ve shown how it excels in embedding and classifying security data while delivering high performance on specialized security tasks.  

The DEMIST-2 advantage

DEMIST-2 was built for precision and reliability. Our primary goal was to create a high-performance model capable of tackling complex cybersecurity tasks. Optimizing for efficiency and scalability came second, but it is a natural outcome of our commitment to building a strong, effective solution that is available to security teams working across diverse environments. It is an enormous benefit that DEMIST-2 is orders of magnitude smaller than many general-purpose models. However, and much more importantly, it significantly outperforms models in its capabilities and accuracy on security tasks.  

Finding a product that fits into an environment’s unique constraints used to mean that some teams had to settle for less powerful or less performant products. With DEMIST-2, data can remain local to the environment, is entirely separate from the data of other customers, and can even operate in environments without network connectivity. The size of our model allows for flexible deployment options while at the same time providing measurable performance advantages for security-related tasks.  

As security threats continue to evolve, we believe that purpose-built AI systems like DEMIST-2 will be essential tools for defenders, combining the power of modern language modeling with the specificity and reliability that builds trust and partnership between security practitioners and AI systems.

Conclusion

DEMIST-2 has additional architectural and deployment updates that improve performance and stability. These innovations contribute to our ability to minimize model size and memory constraints and reflect our dedication to meeting the data handling and privacy needs of security environments. In addition, these choices reflect our dedication to responsible AI practices.

DEMIST-2 is available in Darktrace 6.3, along with a new DIGEST model that uses GNNs and RNNs to score and prioritize threats with expert-level precision.

[related-resource]

Want more details?

Read the full research paper to explore how DEMIST-2 was built, trained, and optimized to meet the unique challenges of cybersecurity

Inside the SOC
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
Margaret Cunningham, PhD
Director, Security & AI Strategy, Field CISO

More in this series

No items found.

Blog

/

/

April 24, 2025

The Importance of NDR in Resilient XDR

picture of hands typing on laptop Default blog imageDefault blog image

As threat actors become more adept at targeting and disabling EDR agents, relying solely on endpoint detection leaves critical blind spots.

Network detection and response (NDR) offers the visibility and resilience needed to catch what EDR can’t especially in environments with unmanaged devices or advanced threats that evade local controls.

This blog explores how threat actors can disable or bypass EDR-based XDR solutions and demonstrates how Darktrace’s approach to NDR closes the resulting security gaps with Self-Learning AI that enables autonomous, real-time detection and response.

Threat actors see local security agents as targets

Recent research by security firms has highlighted ‘EDR killers’: tools that deliberately target EDR agents to disable or damage them. These include the known malicious tool EDRKillShifter, the open source EDRSilencer, EDRSandblast and variants of Terminator, and even the legitimate business application HRSword.

The attack surface of any endpoint agent is inevitably large, whether the software is challenged directly, by contesting its local visibility and access mechanisms, or by targeting the Operating System it relies upon. Additionally, threat actors can readily access and analyze EDR tools, and due to their uniformity across environments an exploit proven in a lab setting will likely succeed elsewhere.

Sophos have performed deep research into the EDRShiftKiller tool, which ESET have separately shown became accessible to multiple threat actor groups. Cisco Talos have reported via TheRegister observing significant success rates when an EDR kill was attempted by ransomware actors.

With the local EDR agent silently disabled or evaded, how will the threat be discovered?

What are the limitations of relying solely on EDR?

Cyber attackers will inevitably break through boundary defences, through innovation or trickery or exploiting zero-days. Preventive measures can reduce but not completely stop this. The attackers will always then want to expand beyond their initial access point to achieve persistence and discover and reach high value targets within the business. This is the primary domain of network activity monitoring and NDR, which includes responsibility for securing the many devices that cannot run endpoint agents.

In the insights from a CISA Red Team assessment of a US CNI organization, the Red Team was able to maintain access over the course of months and achieve their target outcomes. The top lesson learned in the report was:

“The assessed organization had insufficient technical controls to prevent and detect malicious activity. The organization relied too heavily on host-based endpoint detection and response (EDR) solutions and did not implement sufficient network layer protections.”

This proves that partial, isolated viewpoints are not sufficient to track and analyze what is fundamentally a connected problem – and without the added visibility and detection capabilities of NDR, any downstream SIEM or MDR services also still have nothing to work with.

Why is network detection & response (NDR) critical?

An effective NDR finds threats that disable or can’t be seen by local security agents and generally operates out-of-band, acquiring data from infrastructure such as traffic mirroring from physical or virtual switches. This means that the security system is extremely inaccessible to a threat actor at any stage.

An advanced NDR such as Darktrace / NETWORK is fully capable of detecting even high-end novel and unknown threats.

Detecting exploitation of Ivanti CS/PS with Darktrace / NETWORK

On January 9th 2025, two new vulnerabilities were disclosed in Ivanti Connect Secure and Policy Secure appliances that were under malicious exploitation. Perimeter devices, like Ivanti VPNs, are designed to keep threat actors out of a network, so it's quite serious when these devices are vulnerable.

An NDR solution is critical because it provides network-wide visibility for detecting lateral movement and threats that an EDR might miss, such as identifying command and control sessions (C2) and data exfiltration, even when hidden within encrypted traffic and which an EDR alone may not detect.

Darktrace initially detected suspicious activity connected with the exploitation of CVE-2025-0282 on December 29, 2024 – 11 days before the public disclosure of the vulnerability, this early detection highlights the benefits of an anomaly-based network detection method.

Throughout the campaign and based on the network telemetry available to Darktrace, a wide range of malicious activities were identified, including the malicious use of administrative credentials, the download of suspicious files, and network scanning in the cases investigated.

Darktrace / NETWORK’s autonomous response capabilities played a critical role in containment by autonomously blocking suspicious connections and enforcing normal behavior patterns. At the same time, Darktrace Cyber AI Analyst™ automatically investigated and correlated the anomalous activity into cohesive incidents, revealing the full scope of the compromise.

This case highlights the importance of real-time, AI-driven network monitoring to detect and disrupt stealthy post-exploitation techniques targeting unmanaged or unprotected systems.

Unlocking adaptive protection for evolving cyber risks

Darktrace / NETWORK uses unique AI engines that learn what is normal behavior for an organization’s entire network, continuously analyzing, mapping and modeling every connection to create a full picture of your devices, identities, connections, and potential attack paths.

With its ability to uncover previously unknown threats as well as detect known threats using signatures and threat intelligence, Darktrace is an essential layer of the security stack. Darktrace has helped secure customers against attacks including 2024 threat actor campaigns against Fortinet’s FortiManager , Palo Alto firewall devices, and more.  

Stay tuned for part II of this series which dives deeper into the differences between NDR types.

Credit to Nathaniel Jones VP, Security & AI Strategy, FCISO & Ashanka Iddya, Senior Director of Product Marketing for their contribution to this blog.

Continue reading
About the author
Nathaniel Jones
VP, Security & AI Strategy, Field CISO

Blog

/

/

April 22, 2025

Obfuscation Overdrive: Next-Gen Cryptojacking with Layers

man looking at multiple computer screensDefault blog imageDefault blog image

Out of all the services honeypotted by Darktrace, Docker is the most commonly attacked, with new strains of malware emerging daily. This blog will analyze a novel malware campaign with a unique obfuscation technique and a new cryptojacking technique.

What is obfuscation?

Obfuscation is a common technique employed by threat actors to prevent signature-based detection of their code, and to make analysis more difficult. This novel campaign uses an interesting technique of obfuscating its payload.

Docker image analysis

The attack begins with a request to launch a container from Docker Hub, specifically the kazutod/tene:ten image. Using Docker Hub’s layer viewer, an analyst can quickly identify what the container is designed to do. In this case, the container is designed to run the ten.py script which is built into itself.

 Docker Hub Image Layers, referencing the script ten.py.
Figure 1: Docker Hub Image Layers, referencing the script ten.py.

To gain more information on the Python file, Docker’s built in tooling can be used to download the image (docker pull kazutod/tene:ten) and then save it into a format that is easier to work with (docker image save kazutod/tene:ten -o tene.tar). It can then be extracted as a regular tar file for further investigation.

Extraction of the resulting tar file.
Figure 2: Extraction of the resulting tar file.

The Docker image uses the OCI format, which is a little different to a regular file system. Instead of having a static folder of files, the image consists of layers. Indeed, when running the file command over the sha256 directory, each layer is shown as a tar file, along with a JSON metadata file.

Output of the file command over the sha256 directory.
Figure 3: Output of the file command over the sha256 directory.

As the detailed layers are not necessary for analysis, a single command can be used to extract all of them into a single directory, recreating what the container file system would look like:

find blobs/sha256 -type f -exec sh -c 'file "{}" | grep -q "tar archive" && tar -xf "{}" -C root_dir' \;

Result of running the command above.
Figure 4: Result of running the command above.

The find command can then be used to quickly locate where the ten.py script is.

find root_dir -name ten.py

root_dir/app/ten.py

Details of the above ten.py script.
Figure 5: Details of the above ten.py script.

This may look complicated at first glance, however after breaking it down, it is fairly simple. The script defines a lambda function (effectively a variable that contains executable code) and runs zlib decompress on the output of base64 decode, which is run on the reversed input. The script then runs the lambda function with an input of the base64 string, and then passes it to exec, which runs the decoded string as Python code.

To help illustrate this, the code can be cleaned up to this simplified function:

def decode(input):
   reversed = input[::-1]

   decoded = base64.decode(reversed)
   decompressed = zlib.decompress(decoded)
   return decompressed

decoded_string = decode(the_big_text_blob)
exec(decoded_string) # run the decoded string

This can then be set up as a recipe in Cyberchef, an online tool for data manipulation, to decode it.

Use of Cyberchef to decode the ten.py script.
Figure 6: Use of Cyberchef to decode the ten.py script.

The decoded payload calls the decode function again and puts the output into exec. Copy and pasting the new payload into the input shows that it does this another time. Instead of copy-pasting the output into the input all day, a quick script can be used to decode this.

The script below uses the decode function from earlier in order to decode the base64 data and then uses some simple string manipulation to get to the next payload. The script will run this over and over until something interesting happens.

# Decode the initial base64

decoded = decode(initial)
# Remove the first 11 characters and last 3

# so we just have the next base64 string

clamped = decoded[11:-3]

for i in range(1, 100):
   # Decode the new payload

   decoded = decode(clamped)
   # Print it with the current step so we

   # can see what’s going on

   print(f"Step {i}")

   print(decoded)
   # Fetch the next base64 string from the

   # output, so the next loop iteration will

   # decode it

   clamped = decoded[11:-3]

Result of the 63rd iteration of this script.
Figure 7: Result of the 63rd iteration of this script.

After 63 iterations, the script returns actual code, accompanied by an error from the decode function as a stopping condition was never defined. It not clear what the attacker’s motive to perform so many layers of obfuscation was, as one round of obfuscation versus several likely would not make any meaningful difference to bypassing signature analysis. It’s possible this is an attempt to stop analysts or other hackers from reverse engineering the code. However,  it took a matter of minutes to thwart their efforts.

Cryptojacking 2.0?

Cleaned up version of the de-obfuscated code.
Figure 8: Cleaned up version of the de-obfuscated code.

The cleaned up code indicates that the malware attempts to set up a connection to teneo[.]pro, which appears to belong to a Web3 startup company.

Teneo appears to be a legitimate company, with Crunchbase reporting that they have raised USD 3 million as part of their seed round [1]. Their service allows users to join a decentralized network, to “make sure their data benefits you” [2]. Practically, their node functions as a distributed social media scraper. In exchange for doing so, users are rewarded with “Teneo Points”, which are a private crypto token.

The malware script simply connects to the websocket and sends keep-alive pings in order to gain more points from Teneo and does not do any actual scraping. Based on the website, most of the rewards are gated behind the number of heartbeats performed, which is likely why this works [2].

Checking out the attacker’s dockerhub profile, this sort of attack seems to be their modus operandi. The most recent container runs an instance of the nexus network client, which is a project to perform distributed zero-knowledge compute tasks in exchange for cryptocurrency.

Typically, traditional cryptojacking attacks rely on using XMRig to directly mine cryptocurrency, however as XMRig is highly detected, attackers are shifting to alternative methods of generating crypto. Whether this is more profitable remains to be seen. There is not currently an easy way to determine the earnings of the attackers due to the more “closed” nature of the private tokens. Translating a user ID to a wallet address does not appear to be possible, and there is limited public information about the tokens themselves. For example, the Teneo token is listed as “preview only” on CoinGecko, with no price information available.

Conclusion

This blog explores an example of Python obfuscation and how to unravel it. Obfuscation remains a ubiquitous technique employed by the majority of malware to aid in detection/defense evasion and being able to de-obfuscate code is an important skill for analysts to possess.

We have also seen this new avenue of cryptominers being deployed, demonstrating that attackers’ techniques are still evolving - even tried and tested fields. The illegitimate use of legitimate tools to obtain rewards is an increasingly common vector. For example,  as has been previously documented, 9hits has been used maliciously to earn rewards for the attack in a similar fashion.

Docker remains a highly targeted service, and system administrators need to take steps to ensure it is secure. In general, Docker should never be exposed to the wider internet unless absolutely necessary, and if it is necessary both authentication and firewalling should be employed to ensure only authorized users are able to access the service. Attacks happen every minute, and even leaving the service open for a short period of time may result in a serious compromise.

References

1. https://www.crunchbase.com/funding_round/teneo-protocol-seed--a8ff2ad4

2. https://teneo.pro/

Continue reading
About the author
Nate Bill
Threat Researcher
Your data. Our AI.
Elevate your network security with Darktrace AI