ブログ
/
AI
/
August 5, 2021

Detecting a Cobalt Strike Attack With Darktrace AI

See how Darktrace AI was able to detect Cobalt Strike attacks by identifying anomalous connections and performing automated network reconnaissance.
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
Brianna Luong (Leddy)
Sr. Technical Alliances Manager
Default blog image
05
Aug 2021

Since its release in 2012, Cobalt Strike has become a popular platform for red teams and ethical hackers. Robust and reliable software combined with innovative features such as DNS tunnelling, lateral movement tools for privilege escalation, and PowerShell support, have made it a desirable option for organizations wanting to test their own cyber defenses. As the framework was previously only available with a commercial license, it gave security teams a distinct advantage over threat actors when preparing for attacks.

That all changed in late 2020, when a GitHub repository appeared hosting a decompiled version of the framework. Users claimed that the leaked platform did indeed function similarly, if not identically, to the commercial version, and even included a commented-out licensing check. This suddenly made the software readily available, and highly appealing for cyber-criminals: rather than requiring a paper trail and licensing, its source code was freely available for customization and use in offensive campaigns.

With sophisticated capabilities of subtle command and control (C2), privilege escalation, and lateral movement, the tools have become a favorite for ransomware gangs. Even prior to the reporting of the leaked version, 66% of ransomware attacks were found to use Cobalt Strike.

Overview of a Cobalt Strike attack

Cobalt Strike has distinctive TTPs (tools, techniques and procedures) and evasive features for each stage of the attack.

Figure 1: Cyber kill chain with Cobalt Strike

Initial compromise can be achieved with a native module for modifying emails. This includes the insertion of malicious links into existing emails or the creation of convincing spear phishing emails.

The initial payload is intentionally lightweight and can be delivered from cheaply hosted infrastructure. The smaller file size is easier to obfuscate and can be implemented in several ways, including injection into libraries or trusted processes, or creating a series of persistence mechanisms (such as turning off anti-virus prior to downloading the full payload). As such, it is remarkably difficult to detect with blocking rules or signatures.

Network reconnaissance can be done through a variety of subtle methods, using commonly used protocols such as DNS and DCE-RPC to interrogate the network. These services are frequently used in legitimate operations, so it is challenging to apply sufficiently strict controls to prevent this stage of the attack.

Lateral movement and privilege escalation are easily accessible with pre-packaged versions of common attack tools such as Mimikatz. They can interrogate an Active Directory (AD) or steal credentials, while also using SMB pipes for peer-to-peer C2. There is little space for perimeter-based security controls to monitor and restrict these abuses, even if sufficiently granular controls could be imposed.

Payload execution is a straightforward matter as Cobalt Strike beacon allows the delivery of effectively arbitrary payloads, including portability for ransomware. As the previous evasive steps can afford the attacker privileged credentials, the deployment of such payloads could look like non-threatening administrative behavior.

AI detections

Initial compromise

Cobalt Strike has utilities for creating spear phishing documents. As email remains a prolific source of perimeter breaches, threat actors will frequently implant the tool through phishes.

One such example was detected by Darktrace’s AI at Canadian manufacturer in June 2021. The compromise started when an end user appeared to open a phishing document, evidenced by connections to Adobe and VeriSign shortly prior to an HTTP connection to a rare external IP address.

A packet capture of the anomalous connection revealed the creation of an object using a base64 encoded string – a common obfuscation technique. If the customer had been using Darktrace/Email, the threat would have been nullified before it hit the mailbox.

Shortly after the HTTP connection, Darktrace identified unusual use of SSL, which appears to have been leveraged to upgrade to HTTPS using self-signed certificates. The endpoint served an executable, which was later confirmed as a Cobalt Strike beacon based on open-source intelligence (OSINT). Such beacons are supported by the framework, with a variety of common C2 protocols available to the attacker.

Figure 2: Event log for ‘Patient Zero’ of a Sodinokibi infection

Darktrace’s detection was based on the anomalous nature of the connection (suspicious violations of standard SSL protocols) and not a pre-defined rule. The initial compromise was detected in a matter of minutes.

Network reconnaissance

In another example at a Swiss telecommunications company in April 2021, Darktrace alerted the security team that a device – normally used for data collection – was engaging in suspicious lateral movement activity.

The host was abusing privileged credentials to perform AD reconnaissance and SMB enumeration. The alert then prompted a broader investigation, revealing that multiple devices, including domain controllers, were compromised with IoCs related to Cobalt Strike.

Thanks to Darktrace’s deep understanding of the business and recognition that this behavior was anomalous, the security team were able to remediate the infection before file encryption or large data exfiltration had occurred.

Privilege escalation and ransomware deployment

In a ransomware attack against a South African insurance company in May 2021, where a phishing email resulted in the deployment of ransomware, Darktrace first identified the creation of new administrative credentials. The devices which used the credentials were then seen making anomalous connections to various C2 endpoints associated with Cobalt Strike beacons.

Darktrace enabled the rapid identification of compromised hosts, which in turn allowed for a faster remediation and mitigated fears of a resurgent infection.

Cyber AI Analyst performed a machine-speed investigation of the activity, and automatically produced a report highlighting unusual connections on TCP port 4444 as well as other mail related ports. Port 4444 is the default port for Metasploit, another hacking platform which is often seen in conjunction with Cobalt Strike beacon. It then presented the human analysts with a full list of compromised hosts.

Figure 3: Cyber AI Analyst summary of an affected host using non-standard ports for C2 and subsequently scanning the network

Cobalt Strike malware

As it appears that a cheaply accessible analog of Cobalt Strike has been leaked, detection of the framework is critical to defend against active attackers. Signatures and rule-based restrictions prove ineffective in this regard, as the framework was designed specifically to evade such tools.

Darktrace offers the capability to detect malicious activity in its earliest stages, to triage at the speed of AI, and to autonomously block the proliferation of active threats.

Thanks to Darktrace analyst Roberto Romeu for his insights on the above threat find.

Learn how Darktrace caught APT41 leveraging Cobalt Strike

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
Brianna Luong (Leddy)
Sr. Technical Alliances Manager

More in this series

No items found.

Blog

/

Email

/

July 24, 2026

Darktrace / EMAIL Expands Behavioral Defense Across Email and Collaboration Workflows

Default blog imageDefault blog image

Email and collaboration tools do more than carry messages. They are where organizations approve payments, share sensitive data, reset credentials, and make thousands of everyday decisions. Increasingly, they are interfaces through which humans direct AI agents in their daily activity. Email, Slack and Teams are high volume, rich with sensitive data, and an easy place to hide malicious activity.

The opportunity isn’t lost on bad actors. Darktrace / EMAIL detected more than 32 million high-confidence phishing emails globally in 2025, and 70% of those messages passed DMARC authentication.  Phishing is increasingly difficult to detect and familiar trust signals alone are not enough. People and security teams need to understand how a message fits the normal behavior of the sender, recipient, and organization. They also need to correlate activity across platforms to spot threats that span multiple channels.

To effectively secure against today’s evolved threats, security teams need to act at two levels: they need to help each employee make a safer decision ‘in the moment’, and they need to understand the wider patterns that may expose the business to risk.

Darktrace is introducing four new capabilities in Darktrace / EMAIL to address both challenges. The new features explain suspicious content more clearly to end users, strengthen the capabilities of Darktrace / Adaptive Human Defense with richer guidance, let organizations define their own patterns for detecting sensitive data in messages, and give security teams a process-level view of risk across email and collaboration workflows.

Darktrace / EMAIL Inbox Analysis highlights risky content within your emails

A warning is more useful when it explains what the user should look at. To help do that, we’ve expanded Darktrace / EMAIL’s Inbox Analysis Add-In to highlight potentially dangerous content within the body of emails that Darktrace / EMAIL flags as potentially suspicious or high risk.  

The add-in can highlight language designed to create urgency, financial references, requests for payment, suspicious links, and content that is unusual for the sender. Each highlighted element includes a pop up that explains why it may be suspicious. Instead of asking an employee to accept a verdict without context, the analysis helps them examine the message and make a more informed decision.

Enhanced Just-In-Time Training Banners in Darktrace / Adaptive Human Defense

Enhanced Just-In-Time Training Banners build on the same principle. The banners now include a contextual header, actionable advice, and specific detection context. This gives employees more useful guidance at the point of risk without adding unnecessary information or cognitive load.

Together, the capabilities help turn a warning into a short learning moment. Employees can see what looks unusual, understand what action to take, and build their judgment.

Custom Sensitive Data Detection in Darktrace / EMAIL - Data Loss Prevention

Sensitive data is different for every business. Standard categories such as payment card details or government identifiers matter, but organizations also have their own customer codes, project names, research formats, account structures, and internal identifiers.

Custom Sensitive Data Detection in Darktrace / EMAIL - Data Loss Prevention allows administrators to write custom expressions for the data their organization needs to protect. Matched content can trigger existing model actions and data loss prevention (DLP) workflows, extending Darktrace's DLP capabilities.

This extends data loss detection beyond a fixed library of common data types. Security teams can apply controls to information that is sensitive in the context of their own organization and adapt those controls as the business changes.

Introducing Email and Collaboration Workflow Risk Posture Dashboards

Some of the most important risks are not isolated events. They are repeated ways of working that create an opening for error, misuse, or attack. For example, a payment request may be one suspicious message, but a recurring approval workflow that relies on weak verification is a business process risk.

The new Email and Collaboration Workflow Risk Posture Dashboard analyzes email and collaboration data across Email, Microsoft Teams, Slack and Zoom to provide a process-level view of risk in the organization. These may include financial authorization workflows, sensitive data sharing patterns, and activity that could expose credentials.

The dashboard brings these patterns into a view and provides actionable recommendations. This helps security teams determine where to investigate or strengthen controls, where ownership needs to be clarified, and where the business may need to change a risky process. It gives CISOs a clearer view of how human and communication risk is embedded in everyday operations, not only where individual alerts occur.

Behavior connects the individual decision to the wider risk

These capabilities build on Darktrace’s unique behavioral approach to security. We use Adaptive AI to learn how people and AI normally behave within an organization, creating the context needed to recognize when activity changes.

Within the Darktrace Behavioral Defense Platform, Darktrace / EMAIL helps protect people against phishing, account takeover, data exfiltration, and human risk across email and collaboration tools. The new capabilities extend that protection in both directions. They give employees clearer context for the decision in front of them, while giving security leaders a broader view of the workflows and behavior that create risk across the organization.

The result is not simply more alerts. It is a better understanding of why something is risky, what action to take, and where the organization can reduce risk before a familiar process becomes an easy route for an attacker.

[related-resource]

Continue reading
About the author
Carlos Gray
Senior Product Marketing Manager, Email

Blog

/

AI

/

July 24, 2026

When Guardrails Break: Why Securing AI Requires Behavioral Detection and Autonomous Containment

Default blog imageDefault blog image

Bottom line up front: Governance, guardrails, identity controls, and secure development are necessary to secure AI, but they are not sufficient. AI systems are probabilistic, adaptive, and non-deterministic. Therefore, organizations need two additional layers of security:

  1. Behavioral-based detection that can identify when AI begins to act outside its intended purpose; and  
  2. Surgical, explainable autonomous containment that can stop risky activity before it causes material damage.  

That capability depends on multiple specialized AI models working together, not one LLM making every decision.

Organizations are embedding AI into development, business operations, and security workflows faster than most security programs can adapt. The risk is no longer limited to the model. It extends across prompts, data, identities, agents, memory, APIs, tools, permissions, and the trust relationships connecting them.

In my recent blog, Securing AI: Analysis of the Complete Security Stack with Governance and Controls, I outlined a defense-in-depth strategy spanning governance, identity, data security, secure development, runtime detection, autonomous containment, and recovery. The most urgent requirement across that architecture is the ability to understand how AI behaves in practice and contain it when that behavior becomes risky.  

Why non-deterministic systems require behavioral-based detection

Traditional controls remain foundational. Organizations need least privilege, strong identity controls, secure-by-design architecture, data governance, AI inventories, guardrails, testing, and clear boundaries on autonomy.

But deterministic controls, which assume predictable and repeatable behavior, cannot fully secure non-deterministic systems, where the same input may not always produce the same outcome.

AI agents can interpret the same instruction differently, chain individually authorized actions into an unsafe outcome, or pursue a legitimate goal through a method the organization did not anticipate. One of the most recent examples of this is the incident that OpenAI and Hugging Face jointly disclosed, where an autonomous agent escaped its intended testing boundaries and compromised Hugging Face infrastructure.  

An agent may have permission to access data and invoke a tool, but that does not mean every use of that access is appropriate. It is not enough to know whether an action is allowed. Organizations need to know whether it makes sense.

  • Is this normal for this agent?  
  • Is it acting within its intended purpose?  
  • Is it accessing unusual data, invoking an unexpected tool, or beginning to drift?  
  • Do a series of ordinary-looking actions become risky when viewed together?

Behavioral-based detection specific to an environment or organization with an understanding of context and risk enables provides the needed detection engineering for AI systems. It learns normal activity across people, systems, data, devices, and AI agents, then identifies deviations and evaluates their risk, intent, and context. This enables detection of misuse, abuse, compromise, manipulation, and unintended behavior even when no known attack signature or explicit policy violation exists.

Why accuracy is the foundation for SOC optimization

AI will only improve the SOC if it produces accurate, explainable, and actionable outcomes.

If analysts must manually validate every AI-generated finding because they cannot understand the evidence or confidence behind it, automation has not reduced workload. It has moved the workload. False positives increase fatigue. False negatives cause the most risk and damage to organizations. Inaccurate autonomous actions can disrupt critical operations.

Accuracy is therefore more than a model-performance metric. It is the prerequisite for analyst trust, SOC optimization, and safe autonomous response.

That accuracy is unlikely to come from one model.

Generative AI is valuable for natural-language analysis, summarization, and human interaction. But an LLM should not be the sole analytical engine for behavioral-based detection, investigation, risk assessment, and containment. Interpretability and consistency are required for high-consequence security decisions.

A stronger architecture uses multiple specialized AI systems collaboratively:  

  • Behavioral models can establish normal activity.  
  • Unsupervised learning can identify novel anomalies.  
  • Graph analysis can evaluate relationships among agents, identities, systems, and tools.  
  • Other models can correlate events, investigate competing hypotheses, and assess risk.  
  • Semantic models can analyze language where behavior-based language analysis is needed but this can be used in tandem with vector embeddings, graph neural networks, and a variety of other AI systems.

Each model contributes a different analytical perspective. Their outputs can corroborate one another, improving accuracy and creating a more reliable basis for response. The objective is not one model operating as an oracle. It is layered, adaptive intelligence designed to produce decisions the SOC can understand and trust.

Autonomous containment is required to secure autonomous systems

Many SOCs remain hesitant to trust LLM-based agents with autonomous containment. That concern is reasonable. A poorly selected response can isolate the wrong asset, stop a critical workflow, block a legitimate identity, or create more operational damage than the original incident.

But relying exclusively on human response is also not viable.

AI systems can operate at machine speed. They can expose sensitive data, execute workflows, modify records, call tools, or propagate actions across connected systems before an analyst can investigate and intervene. The behavior may be unintentional, the result of an agent optimizing toward a goal, or caused by misuse, compromise, prompt injection, or offensive AI.

Intent affects the investigation. It does not change the need to stop the damage.

Organizations need autonomous response, but it must be surgical and explainable. The objective is not to shut down an entire agent, user, application, or business process whenever an anomaly occurs. It is to interrupt the specific risky behavior: block an unusual connection, constrain a tool call, stop an abnormal data transfer, or temporarily limit an agent when it is performing anomalous, risky activity.  

That buys humans time. It stops the spread, limits damage, and allows the SOC to investigate without unnecessarily disrupting the business.

Layered, Adaptive AI provides a path forward

Darktrace has spent more than a decade researching and operationalizing layered, behavioral, Adaptive AI that learns a specific organization rather than relying only on historic attacks or predefined signatures.

The approach is designed to understand normal behavior, identify anomalous activity, assess its risk, correlate related events, autonomously investigate, and, when necessary, apply targeted containment while normal operations continue.

That sequence matters. Autonomous response cannot simply be added to the end of an LLM workflow. Trusted containment depends on broad visibility, continuous behavioral understanding, multiple analytical techniques, risk and context evaluation, autonomous investigation, explainability, and precise response actions.

This represents a more responsible model for security autonomy: not automation for its own sake, but controlled autonomy built to improve security outcomes and protect business operations.

Security must enable AI adoption

The answer for security teams is not to block AI. Organizations are adopting it to improve productivity, accelerate development, and create new business value.

But innovation without behavioral detection and autonomous containment is not sustainable.

Organizations should continue investing in governance, identity, least privilege, data security, secure MLOps, guardrails, testing, evaluation, validation, verification, kill switches, rollback, and forensic readiness. At the same time, they cannot wait for every governance program to mature before addressing runtime risk.

Behavioral-based detection and autonomous containment provide an immediate layer of resilience. They allow organizations to detect exploitation and risky AI behavior they did not anticipate, contain it at machine speed, and preserve human control over broader remediation.

The future of AI security will not be defined by a single model making every decision. It will be defined by multiple specialized AI systems working collaboratively, with sufficient accuracy, transparency, and context to support trusted autonomous action.

Surgical, explainable autonomous containment is no longer a future capability. It is a requirement for scaling AI securely today.

Learn how to build a defense-in-depth strategy for securing AI at scale in our talk at Black Hat on August 5 at 3:15 PM.  

[related-resource]

Continue reading
About the author
あなたのデータ × DarktraceのAI
唯一無二のDarktrace AIで、ネットワークセキュリティを次の次元へ