Blog
/
/
April 15, 2021

AI Neutralizes Hafnium Cyber Attack in December 2020

Protect your business from cyber attacks with AI technology. Learn how Darktrace neutralized the Hafnium attack against Exchange servers in December 2020.
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
Max Heinemeyer
Global Field CISO
Default blog imageDefault blog imageDefault blog imageDefault blog imageDefault blog imageDefault blog image
15
Apr 2021

In early December 2020, Darktrace AI autonomously detected and investigated a sophisticated cyber-attack that targeted a customer’s Exchange server. On March 2, 2021, Microsoft disclosed an ongoing campaign by the Hafnium threat actor group leveraging Exchange server zero-days.

Based on similarities in techniques, tools and procedures (TTPs) observed, Darktrace has now assessed with high confidence that the attack in December was the work of the Hafnium group. Although it is not possible to determine whether this attack leveraged the same Exchange zero-days as reported by Microsoft, the finding suggests that Hafnium’s campaign was active several months earlier than assumed.

As a result, organizations may want to go back as far as early December 2020 to check security logs and tools for signs of initial intrusion into their Internet-facing Exchange servers.

As Darktrace does not rely on rules or signatures, it doesn’t require a constant cloud connection. Most customers therefore operate our technology themselves, and we don’t centrally monitor their detections.

At the time of detection in December, this was one of many uncategorized, sophisticated intrusions that affected only a single customer, and was not indicative of a broader campaign.

This means that while we protect our customers from individual intrusions, we are not in a position to do global campaign tracking like other companies which focus primarily on threat intelligence and threat actor tracking.

In this blog, we will analyze the attack to aid organizations in their ongoing investigations, and to raise awareness that the Hafnium campaign may have been active for longer than previously disclosed.

Overview of the Exchange attack

The intrusion was detected at an organization in the critical national infrastructure sector in South Asia. One hypothesis is that the Hafnium group was testing out and refining its TTPs, potentially including the Exchange server exploit, before running a broad-scale campaign against Western organizations in early 2021.

The threat actor used many of the same techniques that were observed in the later Hafnium attacks, including the deployment of the low-activity China Chopper web shell, quickly followed by post-exploitation activity – attempting to move laterally and spread to critical devices in the network.

The following analysis demonstrates how Darktrace’s Enterprise Immune System detected the malicious activity, how Cyber AI Analyst automatically investigated on the incident and surfaced the alert as a top priority, and how Darktrace RESPOND (formerly known as 'Antigena') would have responded autonomously to shut down the attack, had it been in active mode.

All the activity took place in early December 2020, almost three months before Microsoft released information about the Hafnium campaign.

Figure 1: Timeline of the attack from early December 2020

Initial compromise

Unfortunately, the victim organization did not keep any logs or forensic artefacts from their Exchange server in December 2020, which would have allowed Darktrace to ascertain the exploit of the zero-day. However, there is circumstantial evidence suggesting that these Exchange server vulnerabilities were abused.

Darktrace observed no signs of compromise or change in behavior from the Internet-facing Exchange server – no prior internal admin connections, no broad-scale brute-force attempts, no account takeovers, no malware copied to the server via internal channels – until all of a sudden, it began to scan the internal network.

While this is not conclusive evidence that no other avenue of initial intrusion was present, the change in behavior on an administrative level points to a complete takeover of the Exchange server, rather than the compromise of a single Outlook Web Application account.

To conduct a network scan from an Exchange server, a highly privileged, operating SYSTEM-level account is required. The patch level of the Exchange server at the time of compromise appears to have been up-to-date, at least not offering a threat actor the ability to target a known vulnerability to instantly get SYSTEM-level privileges.

For this reason, Darktrace has inferred that the Exchange server zero-days that became public in early March 2021 were possibly being used in this attack observed in early December 2020.

Internal reconnaissance

As soon as the attackers gained access via the web shell, they used the Exchange server to scan all IPs in a single subnet on ports 80, 135, 445, 8080.

This particular Exchange server had never made such a large number of new failed internal connections to that specific subnet on those key ports. As a result, Darktrace instantly alerted on the anomalous behavior, which was indicative of a network scan.

Autonomous Response

Darktrace RESPOND was in passive mode in the environment, so was not able to take action. In active mode, it would have responded by enforcing the previously learned, normal ‘pattern of life’ of the Exchange server – allowing the server to continue normal business operations (sending and receiving emails) but preventing the network scan and any subsequent activity. These actions would have been carried out via various integrations with the customer’s existing security stack, including Firewalls and Network Access Controls.

Specifically, when the network scanning started, the ‘Antigena Network Scan Block’ was triggered. This means that for several hours, Darktrace RESPOND (Antigena) would have blocked any new outgoing connections from the Exchange server to the scanned subnet on port 80, 135, 445, or 8080, preventing the infected Exchange server from conducting network scanning.

As a result, the attackers would not have been able to conclude anything from their reconnaissance — all their scanning would have returned closed ports. At this point, they would need to stop their attack or resort to other means, likely triggering further detections and further Autonomous Response.

The network scan was the first step touching the internal network. This is therefore a clear case of how Darktrace RESPOND can intercept an attack in seconds, acting at the earliest possible evidence of the intrusion.

Lateral movement

Less than an hour after the internal network scan, the compromised Exchange server was observed writing further web shells to other Exchange servers via internal SMB. Darktrace alerted on this as the initially compromised Exchange server had never accessed the other Exchange servers in this fashion over SMB, let alone writing .aspx files to Program Files remotely.

A single click allowed the security team to pivot from the alert into Darktrace’s Advanced Search, revealing further details about the written files. The full file path for the newly deployed web shells was:

Program Files\Microsoft\Exchange Server\V15\FrontEnd\HttpProxy\owa\auth\Current\themes\errorFS.aspx

The attackers thus used internal SMB to compromise further Exchange servers and deploy more web shells, rather than using the Exchange zero-day exploit again to achieve the same goal. The reason for this is clear: exploits can often be unstable, and an adversary would not want to show their hand unnecessarily if it could be avoided.

While the China Chopper web shell has been deployed with many different names in the past, the file path and file name of the actual .aspx web shell bear very close resemblance to the Hafnium campaign details published by Microsoft and others in March 2021.

As threat actors often reuse naming conventions / TTPs in coherent campaigns, it again indicates that this particular attack was in some way part of the broader campaign observed in early 2021.

Further lateral movement

Minutes later, the attacker conducted further lateral movement by making more SMB drive writes to Domain Controllers. This time the attackers did not upload web shells, but malware, in the form of executables and Windows .bat files.

Darktrace alerted the security team as it was extremely unusual for the Exchange server and its peer group to make SMB drive writes to hidden shares to a Domain Controller, particularly using executables and batch files. The activity was presented to the team in the form of a high-confidence alert such as the anonymized example below.

Figure 2: Example graphic of Darktrace detecting unusual connectivity

The batch file was called ‘a.bat’. At this point, the security team could have created a packet capture for the a.bat file in Darktrace with the click of a button, inspecting the content and details of that script at the time of the intrusion.

Darktrace also listed the credentials involved in the activity, providing context into the compromised accounts. This allows an analyst to pivot rapidly around the data and further understand the scope of the intrusion.

Bird’s-eye perspective

In addition to detecting the malicious activity outlined above, Darktrace’s Cyber AI Analyst autonomously summarized the incident and reported on it, outlining the internal reconnaissance and lateral movement activity in a single, cohesive incident.

The organization has several thousand devices covered by Darktrace’s Enterprise Immune System. Nevertheless, over the period of one week, the Hafnium intrusion was in the top five incidents highlighted in Cyber AI Analyst. Even a small or resource-stretched security team, with only a few minutes available per week to review the highest-severity incidents, could have seen and inspected this threat.

Below is a graphic showing a similar Cyber AI Analyst incident created by Darktrace.

Figure 3: A Cyber AI Analyst report showing unusual SMB activity

How to stop a zero-day

Large scale campaigns which target Internet-facing infrastructure and leverage zero-day exploits will continue to occur regularly, and such attacks will always succeed in evading signature-based detection. However, organizations are not helpless against the next high-profile zero-day or supply chain attack.

Detecting the movements of attackers inside a system and responding to contain in-progress threats is possible before IoCs have been provided. The methods of detection outlined above protected the company against this attack in December, and the same techniques will continue to protect the company against unknown threats in the future.

Learn more about how Darktrace AI has stopped Hafnium cyber-attacks and similar threat actors

Darktrace model detections:

  • Device / New or Uncommon WMI Activity
  • Executable Uploaded to DC
  • Compliance / High Priority Compliance Model Breach
  • Compliance / SMB Drive Write
  • Antigena / Network / Insider Threat / Antigena Network Scan Block
  • Device / Network Scan - Low Anomaly Score
  • Unusual Activity / Unusual Internal Connections

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
Max Heinemeyer
Global Field CISO

More in this series

No items found.

Blog

/

/

May 8, 2025

Anomaly-based threat hunting: Darktrace's approach in action

person working on laptopDefault blog imageDefault blog image

What is threat hunting?

Threat hunting in cybersecurity involves proactively and iteratively searching through networks and datasets to detect threats that evade existing automated security solutions. It is an important component of a strong cybersecurity posture.

There are several frameworks that Darktrace analysts use to guide how threat hunting is carried out, some of which are:

  • MITRE Attack
  • Tactics, Techniques, Procedures (TTPs)
  • Diamond Model for Intrusion Analysis
  • Adversary, Infrastructure, Victims, Capabilities
  • Threat Hunt Model – Six Steps
  • Purpose, Scope, Equip, Plan, Execute, Feedback
  • Pyramid of Pain

These frameworks are important in baselining how to run a threat hunt. There are also a combination of different methods that allow defenders diversity– regardless of whether it is a proactive or reactive threat hunt. Some of these are:

  • Hypothesis-based threat hunting
  • Analytics-driven threat hunting
  • Automated/machine learning hunting
  • Indicator of Compromise (IoC) hunting
  • Victim-based threat hunting

Threat hunting with Darktrace

At its core, Darktrace relies on anomaly-based detection methods. It combines various machine learning types that allows it to characterize what constitutes ‘normal’, based on the analysis of many different measures of a device or actor’s behavior. Those types of learning are then curated into what are called models.

Darktrace models leverage anomaly detection and integrate outputs from Darktrace Deep Packet Inspection, telemetry inputs, and additional modules, creating tailored activity detection.

This dynamic understanding allows Darktrace to identify, with a high degree of precision, events or behaviors that are both anomalous and unlikely to be benign.  On top of machine learning models for detection, there is also the ability to change and create models showcasing the tool’s diversity. The Model Editor allows security teams to specify values, priorities, thresholds, and actions they want to detect. That means a team can create custom detection models based on specific use cases or business requirements. Teams can also increase the priority of existing detections based on their own risk assessments to their environment.

This level of dexterity is particularly useful when conducting a threat hunt. As described above, and in previous ‘Inside the SOC’ blogs such a threat hunt can be on a specific threat actor, specific sector, or a  hypothesis-based threat hunt combined with ‘experimenting’ with some of Darktrace’s models.

Conducting a threat hunt in the energy sector with experimental models

In Darktrace’s recent Threat Research report “AI & Cybersecurity: The state of cyber in UK and US energy sectors” Darktrace’s Threat Research team crafted hypothesis-driven threat hunts, building experimental models and investigating existing models to test them and detect malicious activity across Darktrace customers in the energy sector.

For one of the hunts, which hypothesised utilization of PerfectData software and multi-factor authentication (MFA) bypass to compromise user accounts and destruct data, an experimental model was created to detect a Software-as-a-Service (SaaS) user performing activity relating to 'PerfectData Software’, known to allow a threat actor to exfiltrate whole mailboxes as a PST file. Experimental model alerts caused by this anomalous activity were analyzed, in conjunction with existing SaaS and email-related models that would indicate a multi-stage attack in line with the hypothesis.

Whilst hunting, Darktrace researchers found multiple model alerts for this experimental model associated with PerfectData software usage, within energy sector customers, including an oil and gas investment company, as well as other sectors. Upon further investigation, it was also found that in June 2024, a malicious actor had targeted a renewable energy infrastructure provider via a PerfectData Software attack and demonstrated intent to conduct an Operational Technology (OT) attack.

The actor logged into Azure AD from a rare US IP address. They then granted Consent to ‘eM Client’ from the same IP. Shortly after, the actor granted ‘AddServicePrincipal’ via Azure to PerfectData Software. Two days later, the actor created a  new email rule from a London IP to move emails to an RSS Feed Folder, stop processing rules, and mark emails as read. They then accessed mail items in the “\Sent” folder from a malicious IP belonging to anonymization network,  Private Internet Access Virtual Private Network (PIA VPN) [1]. The actor then conducted mass email deletions, deleting multiple instances of emails with subject “[Name] shared "[Company Name] Proposal" With You” from the  “\Sent folder”. The emails’ subject suggests the email likely contains a link to file storage for phishing purposes. The mass deletion likely represented an attempt to obfuscate a potential outbound phishing email campaign.

The Darktrace Model Alert that triggered for the mass deletes of the likely phishing email containing a file storage link.
Figure 1: The Darktrace Model Alert that triggered for the mass deletes of the likely phishing email containing a file storage link.

A month later, the same user was observed downloading mass mLog CSV files related to proprietary and Operational Technology information. In September, three months after the initial attack, another mass download of operational files occurred by this actor, pertaining to operating instructions and measurements, The observed patience and specific file downloads seemingly demonstrated an intent to conduct or research possible OT attack vectors. An attack on OT could have significant impacts including operational downtime, reputational damage, and harm to everyday operations. Darktrace alerted the impacted customer once findings were verified, and subsequent actions were taken by the internal security team to prevent further malicious activity.

Conclusion

Harnessing the power of different tools in a security stack is a key element to cyber defense. The above hypothesis-based threat hunt and custom demonstrated intent to conduct an experimental model creation demonstrates different threat hunting approaches, how Darktrace’s approach can be operationalized, and that proactive threat hunting can be a valuable complement to traditional security controls and is essential for organizations facing increasingly complex threat landscapes.

Credit to Nathaniel Jones (VP, Security & AI Strategy, Field CISO at Darktrace) and Zoe Tilsiter (EMEA Consultancy Lead)

References

  1. https://spur.us/context/191.96.106.219

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

Blog

/

/

May 6, 2025

Combatting the Top Three Sources of Risk in the Cloud

woman working on laptopDefault blog imageDefault blog image

With cloud computing, organizations are storing data like intellectual property, trade secrets, Personally Identifiable Information (PII), proprietary code and statistics, and other sensitive information in the cloud. If this data were to be accessed by malicious actors, it could incur financial loss, reputational damage, legal liabilities, and business disruption.

Last year data breaches in solely public cloud deployments were the most expensive type of data breach, with an average of $5.17 million USD, a 13.1% increase from the year before.

So, as cloud usage continues to grow, the teams in charge of protecting these deployments must understand the associated cybersecurity risks.

What are cloud risks?

Cloud threats come in many forms, with one of the key types consisting of cloud risks. These arise from challenges in implementing and maintaining cloud infrastructure, which can expose the organization to potential damage, loss, and attacks.

There are three major types of cloud risks:

1. Misconfigurations

As organizations struggle with complex cloud environments, misconfiguration is one of the leading causes of cloud security incidents. These risks occur when cloud settings leave gaps between cloud security solutions and expose data and services to unauthorized access. If discovered by a threat actor, a misconfiguration can be exploited to allow infiltration, lateral movement, escalation, and damage.

With the scale and dynamism of cloud infrastructure and the complexity of hybrid and multi-cloud deployments, security teams face a major challenge in exerting the required visibility and control to identify misconfigurations before they are exploited.

Common causes of misconfiguration come from skill shortages, outdated practices, and manual workflows. For example, potential misconfigurations can occur around firewall zones, isolated file systems, and mount systems, which all require specialized skill to set up and diligent monitoring to maintain

2. Identity and Access Management (IAM) failures

IAM has only increased in importance with the rise of cloud computing and remote working. It allows security teams to control which users can and cannot access sensitive data, applications, and other resources.

Cybersecurity professionals ranked IAM skills as the second most important security skill to have, just behind general cloud and application security.

There are four parts to IAM: authentication, authorization, administration, and auditing and reporting. Within these, there are a lot of subcomponents as well, including but not limited to Single Sign-On (SSO), Two-Factor Authentication (2FA), Multi-Factor Authentication (MFA), and Role-Based Access Control (RBAC).

Security teams are faced with the challenge of allowing enough access for employees, contractors, vendors, and partners to complete their jobs while restricting enough to maintain security. They may struggle to track what users are doing across the cloud, apps, and on-premises servers.

When IAM is misconfigured, it increases the attack surface and can leave accounts with access to resources they do not need to perform their intended roles. This type of risk creates the possibility for threat actors or compromised accounts to gain access to sensitive company data and escalate privileges in cloud environments. It can also allow malicious insiders and users who accidentally violate data protection regulations to cause greater damage.

3. Cross-domain threats

The complexity of hybrid and cloud environments can be exploited by attacks that cross multiple domains, such as traditional network environments, identity systems, SaaS platforms, and cloud environments. These attacks are difficult to detect and mitigate, especially when a security posture is siloed or fragmented.  

Some attack types inherently involve multiple domains, like lateral movement and supply chain attacks, which target both on-premises and cloud networks.  

Challenges in securing against cross-domain threats often come from a lack of unified visibility. If a security team does not have unified visibility across the organization’s domains, gaps between various infrastructures and the teams that manage them can leave organizations vulnerable.

Adopting AI cybersecurity tools to reduce cloud risk

For security teams to defend against misconfigurations, IAM failures, and insecure APIs, they require a combination of enhanced visibility into cloud assets and architectures, better automation, and more advanced analytics. These capabilities can be achieved with AI-powered cybersecurity tools.

Such tools use AI and automation to help teams maintain a clear view of all their assets and activities and consistently enforce security policies.

Darktrace / CLOUD is a Cloud Detection and Response (CDR) solution that makes cloud security accessible to all security teams and SOCs by using AI to identify and correct misconfigurations and other cloud risks in public, hybrid, and multi-cloud environments.

It provides real-time, dynamic architectural modeling, which gives SecOps and DevOps teams a unified view of cloud infrastructures to enhance collaboration and reveal possible misconfigurations and other cloud risks. It continuously evaluates architecture changes and monitors real-time activity, providing audit-ready traceability and proactive risk management.

Real-time visibility into cloud assets and architectures built from network, configuration, and identity and access roles. In this unified view, Darktrace / CLOUD reveals possible misconfigurations and risk paths.
Figure 1: Real-time visibility into cloud assets and architectures built from network, configuration, and identity and access roles. In this unified view, Darktrace / CLOUD reveals possible misconfigurations and risk paths.

Darktrace / CLOUD also offers attack path modeling for the cloud. It can identify exposed assets and highlight internal attack paths to get a dynamic view of the riskiest paths across cloud environments, network environments, and between – enabling security teams to prioritize based on unique business risk and address gaps to prevent future attacks.  

Darktrace’s Self-Learning AI ensures continuous cloud resilience, helping teams move from reactive to proactive defense.

[related-resource]

Continue reading
About the author
Pallavi Singh
Product Marketing Manager, OT Security & Compliance
Your data. Our AI.
Elevate your network security with Darktrace AI