April 5, 2023

Understanding Qakbot Infections and Attack Paths

Explore the network-based analysis of Qakbot infections with Darktrace. Learn about the various attack paths used by cybercriminals and Darktrace's response.

Written by

Sam Lister

Specialist Security Researcher

Written by

Connor Mooney

SOC Analyst

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

Sam Lister

Specialist Security Researcher

Written by

Connor Mooney

SOC Analyst

Apr 2023

In an ever-changing threat landscape, security vendors around the world are forced to quickly adapt, react, and respond to known attack vectors and threats. In the face of this, malicious actors are constantly looking for novel ways to gain access to networks. Whether that’s through new exploitations of network vulnerabilities or new delivery methods, attackers and their methods are continually evolving. Although it is valuable for organizations to leverage threat intelligence to keep abreast of known threats to their networks, intelligence alone is not enough to defend against increasingly versatile attackers. Having an autonomous decision maker able to detect and respond to emerging threats, even those employing novel or unknown techniques, is paramount to defend against network compromise.

At the end of January 2023, threat actors began to abuse OneNote attachments to deliver the malware strain, Qakbot, onto users' devices. Widespread adoption of this novel delivery method resulted in a surge in Qakbot infections across Darktrace's customer base between the end of January 2023 and the end of February 2023. Using its Self-Learning AI, Darktrace was able to uncover and respond to these so-called ‘QakNote’ infections as the new trend emerged. Darktrace detected and responded to the threat at multiple stages of the kill chain, preventing damaging and widespread compromise to customer networks.

Qakbot and The Recent Weaponization of OneNote

Qakbot first appeared in 2007 as a banking trojan designed to steal sensitive data such as banking credentials. Since then, Qakbot has evolved into a highly modular, multi-purpose tool, with backdoor, payload delivery, reconnaissance, lateral movement, and data exfiltration capabilities. Although Qakbot's primary delivery method has always been email-based, threat actors have been known to modify their email-based delivery methods of Qakbot in the face of changing circumstances. In the first half of 2022, Microsoft started rolling out versions of Office which block XL4 and VBA macros by default [1]/[2]/[3]. Prior to this change, Qakbot email campaigns typically consisted in the spreading of deceitful emails with Office attachments containing malicious macros. In the face of Microsoft's default blocking of macros, threat actors appeared to cease delivering Qakbot via Office attachments, and shifted to primarily using HTML attachments, through a method known as 'HTML smuggling' [4]/[5]. After the public disclosure [6] of the Follina vulnerability (CVE-2022-30190) in Microsoft Support Diagnostic Tool (MSDT) in May 2022, Qakbot actors were seen capitalizing on the vulnerability to facilitate their email-based delivery of Qakbot payloads [7]/[8]/[9].

Given the inclination of Qakbot actors to adapt their email-based delivery methods, it is no surprise that they were quick to capitalize on the novel OneNote-based delivery method which emerged in December 2022. Since December 2022, threat actors have been seen using OneNote attachments to deliver a variety of malware strains, ranging from Formbook [10] to AsynRAT [11] to Emotet [12]. The abuse of OneNote documents to deliver malware is made possible by the fact that OneNote allows for the embedding of executable file types such as HTA files, CMD files, and BAT files. At the end of January 2023, actors started to leverage OneNote attachments to deliver Qakbot [13]/[14]. The adoption of this novel delivery method by Qakbot actors resulted in a surge in Qakbot infections in the wider threat landscape and across the Darktrace customer base.

Observed Activity Chains

Between January 31 and February 24, 2023, Darktrace observed variations of the following pattern of activity across its customer base:

1. User's device contacts OneNote-related endpoint

2. User's device makes an external GET request with an empty Host header, a target URI whose final segment consists in 5 or 6 digits followed by '.dat', and a User-Agent header referencing either cURL or PowerShell. The GET request is responded to with a DLL file

3. User's device makes SSL connections over ports 443 and 2222 to unusual external endpoints, and makes TCP connections over port 65400 to 23.111.114[.]52

4. User's device makes SSL connections over port 443 to an external host named 'bonsars[.]com' (IP: 194.165.16[.]56) and TCP connections over port 443 to 78.31.67[.]7

5. User’s device makes call to Endpoint Mapper service on internal systems and then connects to the Service Control Manager (SCM)

6. User's device uploads files with algorithmically generated names and ‘.dll’ or ‘.dll.cfg’ file extensions to SMB shares on internal systems

7. User's device makes Service Control requests to the systems to which it uploaded ‘.dll’ and ‘.dll.cfg’ files

Further investigation of these chains of activity revealed that they were parts of Qakbot infections initiated via interactions with malicious OneNote attachments.

Figure 1: Steps of observed QakNote infections.

Delivery Phase

Users' interactions with malicious OneNote attachments, which were evidenced by devices' HTTPS connections to OneNote-related endpoints, such as 'www.onenote[.]com', 'contentsync.onenote[.]com', and 'learningtools.onenote[.]com', resulted in the retrieval of Qakbot DLLs from unusual, external endpoints. In some cases, the user's interaction with the malicious OneNote attachment caused their device to fetch a Qakbot DLL using cURL, whereas, in other cases, it caused their device to download a Qakbot DLL using PowerShell. These different outcomes reflected variations in the contents of the executable files embedded within the weaponized OneNote attachments. In addition to having cURL and PowerShell User-Agent headers, the HTTP requests triggered by interaction with these OneNote attachments had other distinctive features, such as empty host headers and target URIs whose last segment consists in 5 or 6 digits followed by '.dat'.

Figure 2: Model breach highlighting a user’s device making a HTTP GET request to 198.44.140[.]78 with a PowerShell User-Agent header and the target URI ‘/210/184/187737.dat’.

Figure 3: Model breach highlighting a user’s device making a HTTP GET request to 103.214.71[.]45 with a cURL User-Agent header and the target URI ‘/70802.dat’.

Figure 4: Event Log showing a user’s device making a GET request with a cURL User-Agent header to 185.231.205[.]246 after making an SSL connection to contentsync.onenote[.]com.

Figure 5: Event Log showing a user’s device making a GET request with a cURL User-Agent header to 185.231.205[.]246 after making an SSL connection to www.onenote[.]com.

Command and Control Phase

After fetching Qakbot DLLs, users’ devices were observed making numerous SSL connections over ports 443 and 2222 to highly unusual, external endpoints, as well as large volumes of TCP connections over port 65400 to 23.111.114[.]52. These connections represented Qakbot-infected devices communicating with command and control (C2) infrastructure. Qakbot-infected devices were also seen making intermittent connections to legitimate endpoints, such as 'xfinity[.]com', 'yahoo[.]com', 'verisign[.]com', 'oracle[.]com', and 'broadcom[.]com', likely due to Qakbot making connectivity checks.

Figure 6: Event Log showing a user’s device contacting Qakbot C2 infrastructure and making connectivity checks to legitimate domains.

Figure 7: Event Log showing a user’s device contacting Qakbot C2 infrastructure and making connectivity checks to legitimate domains.

Cobalt Strike and VNC Phase

After Qakbot-infected devices established communication with C2 servers, they were observed making SSL connections to the external endpoint, bonsars[.]com, and TCP connections to the external endpoint, 78.31.67[.]7. The SSL connections to bonsars[.]com were C2 connections from Cobalt Strike Beacon, and the TCP connections to 78.31.67[.]7 were C2 connections from Qakbot’s Virtual Network Computing (VNC) module [15]/[16]. The occurrence of these connections indicate that actors leveraged Qakbot infections to drop Cobalt Strike Beacon along with a VNC payload onto infected systems. The deployment of Cobalt Strike and VNC likely provided actors with ‘hands-on-keyboard’ access to the Qakbot-infected systems.

Figure 8: Advanced Search logs showing a user’s device contacting OneNote endpoints, fetching a Qakbot DLL over HTTP, making SSL connections to Qakbot infrastructure and connectivity checks to legitimate domains, and then making SSL connections to the Cobalt Strike endpoint, bonsars[.]com.

Figure 9: Event Log showing a user’s device contacting the Cobalt Strike C2 endpoint, bonsars[.]com, and the VNC C2 endpoint, 78.31.67[.]7, whilst simultaneously contacting the Qakbot C2 endpoint, 47.32.78[.]150.

Lateral Movement Phase

After dropping Cobalt Strike Beacon and a VNC module onto Qakbot-infected systems, actors leveraged their strengthened foothold to connect to the Service Control Manager (SCM) on internal systems in preparation for lateral movement. Before connecting to the SCM, infected systems were seen making calls to the Endpoint Mapper service, likely to identify exposed Microsoft Remote Procedure Call (MSRPC) services on internal systems. The MSRPC service, Service Control Manager (SCM), is known to be abused by Cobalt Strike to create and start services on remote systems. Connections to this service were evidenced by OpenSCManager2 (Opnum: 0x40) and OpenSCManagerW (Opnum: 0xf) calls to the svcctl RPC interface.

Figure 10: Advanced Search logs showing a user’s device contacting the Endpoint Mapper and Service Control Manager (SCM) services on internal systems.

After connecting to the SCM on internal systems, infected devices were seen using SMB to distribute files with ‘.dll’ and ‘.dll.cfg’ extensions to SMB shares. These uploads were followed by CreateWowService (Opnum: 0x3c) calls to the svcctl interface, likely intended to execute the uploaded payloads. The naming conventions of the uploaded files indicate that they were Qakbot payloads.

Figure 11: Advanced Search logs showing a user’s device making Service Control DCE-RPC requests to internal systems after uploading ‘.dll’ and ‘.dll.cfg’ files to them over SMB.

Fortunately, none of the observed QakNote infections escalated further than this. If these infections had escalated, it is likely that they would have resulted in the widespread detonation of additional malicious payloads, such as ransomware.

Darktrace Coverage of QakNote Activity

Figure 1 shows the steps involved in the QakNote infections observed across Darktrace’s customer base. How far attackers got along this chain was in part determined by the following three factors:

Whether the targeted customer had Darktrace/Email
Whether the targeted customer had Darktrace RESPOND/Network
Whether the targeted customer acted in response to Darktrace DETECT/Network alerts

The presence of Darktrace/Email typically stopped QakNote infections from moving past the initial infection stage. The presence of RESPOND/Network significantly slowed down observed activity chains, however, infections left unattended and not mitigated by the security teams were able to progress further along the attack chain.

Darktrace observed varying properties in the QakNote emails detected across the customer base. OneNote attachments were typically detected as either ‘application/octet-stream’ files or as ‘application/x-tar’ files. In some cases, the weaponized OneNote attachment embedded a malicious file, whereas in other cases, the OneNote file embedded a malicious link (typically a ‘.png’ or ‘.gif’ link) instead. In all cases Darktrace observed, QakNote emails used subject lines starting with ‘RE’ or ‘FW’ to manipulating their recipients into thinking that such emails were part of an existing email chain/thread. In some cases, emails impersonated users known to their recipients by including the names of such users in their header-from personal names. In many cases, QakNote emails appear to have originated from likely hijacked email accounts. These are highly successful methods of social engineering often employed by threat actors to exploit a user’s trust in known contacts or services, convincing them to open malicious emails and making it harder for security tools to detect.

The fact that observed QakNote emails used the fake-reply method, were sent from unknown email accounts, and contained attachments with unusual MIME types, caused such emails to breach the following Darktrace/Email models:

Association / Unknown Sender
Attachment / Unknown File
Attachment / Unsolicited Attachment
Attachment / Highly Unusual Mime
Attachment / Unsolicited Anomalous Mime
Attachment / Unusual Mime for Organisation
Unusual / Fake Reply
Unusual / Unusual Header TLD
Unusual / Fake Reply + Unknown Sender
Unusual / Unusual Connection from Unknown
Unusual / Off Topic

QakNote emails impersonating known users also breached the following DETECT & RESPOND/Email models:

Unusual / Unrelated Personal Name Address
Spoof / Basic Known Entity Similarities
Spoof / Internal User Similarities
Spoof / External User Similarities
Spoof / Internal User Similarities + Unrelated Personal Name Address
Spoof / External User Similarities + Unrelated Personal Name Address
Spoof / Internal User Similarities + Unknown File
Spoof / External User Similarities + Fake Reply
Spoof / Possible User Spoof from New Address - Enhanced Internal Similarities
Spoof / Whale

The actions taken by Darktrace on the observed emails is ultimately determined by Darktrace/Email models are breached. Those emails which did not breach Spoofing models (due to lack of impersonation indicators) received the ‘Convert Attachment’ action. This action converts suspicious attachments into neutralized PDFs, in this case successfully unweaponizing the malicious OneNote attachments. QakNote emails which did breach Spoofing models (due to the presence of impersonation indicators) received the strongest possible action, ‘Hold Message’. This action prevents suspicious emails from reaching the recipients’ mailbox.

Figure 12: Email log showing a malicious OneNote email (without impersonation indicators) which received a 87% anomaly score, a ‘Move to junk’ action, and a ‘Convert attachment’ actions from Darktrace/Email.

Figure 13: Email log showing a malicious OneNote email (with impersonation indicators) which received an anomaly score of 100% and a ‘Hold message’ action from Darktrace/Email.

Figure 14: Email log showing a malicious OneNote email (with impersonation indicators) which received an anomaly score of 100% and a ‘Hold message’ action from Darktrace/Email.

If threat actors managed to get past the first stage of the QakNote kill chain, likely due to the absence of appropriate email security tools, the execution of the subsequent steps resulted in strong intervention from Darktrace/Network.

Interactions with malicious OneNote attachments caused their devices to fetch a Qakbot DLL from a remote server via HTTP GET requests with an empty Host header and either a cURL or PowerShell User-Agent header. These unusual HTTP behaviors caused the following Darktrace/Network models to breach:

Device / New User Agent
Device / New PowerShell User Agent
Device / New User Agent and New IP
Anomalous Connection / New User Agent to IP Without Hostname
Anomalous Connection / Powershell to Rare External
Anomalous File / Numeric File Download
Anomalous File / EXE from Rare External Location
Anomalous File / New User Agent Followed By Numeric File Download

For customers with RESPOND/Network active, these breaches resulted in the following autonomous actions:

Enforce group pattern of life for 30 minutes
Enforce group pattern of life for 2 hours
Block connections to relevant external endpoints over relevant ports for 2 hours
Block all outgoing traffic for 10 minutes

Figure 15: Event Log showing a user’s device receiving Darktrace RESPOND/Network actions after downloading a Qakbot DLL.

Figure 16: Event Log showing a user’s device receiving Darktrace RESPOND/Network actions after downloading a Qakbot DLL.

Successful, uninterrupted downloads of Qakbot DLLs resulted in connections to Qakbot C2 servers, and subsequently to Cobalt Strike and VNC C2 connections. These C2 activities resulted in breaches of the following DETECT/Network models:

Compromise / Suspicious TLS Beaconing To Rare External
Compromise / Large Number of Suspicious Successful Connections
Compromise / Large Number of Suspicious Failed Connections
Compromise / Sustained SSL or HTTP Increase
Compromise / Sustained TCP Beaconing Activity To Rare Endpoint
Compromise / Beaconing Activity To External Rare
Compromise / Slow Beaconing Activity To External Rare
Anomalous Connection / Multiple Connections to New External TCP Port
Anomalous Connection / Multiple Failed Connections to Rare Endpoint
Device / Initial Breach Chain Compromise

For customers with RESPOND/Network active, these breaches caused RESPOND to autonomously perform the following actions:

Block connections to relevant external endpoints over relevant ports for 1 hour

Figure 17: Event Log showing a user’s device receiving RESPOND/Network actions after contacting the Qakbot C2 endpoint, Cobalt Strike C2 endpoint, bonsars[.]com.

In cases where C2 connections were allowed to continue, actors attempted to move laterally through usage of SMB and Service Control Manager. This lateral movement activity caused the following DETECT/Network models to breach:

Device / Possible SMB/NTLM Reconnaissance
Anomalous Connection / New or Uncommon Service Control

For customers with RESPOND/Network enabled, these breaches caused RESPOND to autonomously perform the following actions:

Block connections to relevant internal endpoints over port 445 for 1 hour

Figure 18: Event Log shows a user’s device receiving RESPOND/Network actions after contacting the Qakbot C2 endpoint, 5.75.205[.]43, and distributing ‘.dll’ and ‘.dll.cfg’ files internally.

The QakNote infections observed across Darktrace’s customer base involved several steps, each of which elicited alerts and autonomous preventative actions from Darktrace. By autonomously investigating the alerts from DETECT, Darktrace’s Cyber AI Analyst was able to connect the distinct steps of observed QakNote infections into single incidents. It then produced incident logs to present in-depth details of the activity it uncovered, provide full visibility for customer security teams.

Figure 19: AI Analyst incident entry showing the steps of a QakNote infection which AI Analyst connected following its autonomous investigations.

Conclusion

Faced with the emerging threat of QakNote infections, Darktrace demonstrated its ability to autonomously detect and respond to arising threats in a constantly evolving threat landscape. The attack chains which Darktrace observed across its customer base involved the delivery of Qakbot via malicious OneNote attachments, the usage of ports 65400 and 2222 for Qakbot C2 communication, the usage of Cobalt Strike Beacon and VNC for ‘hands-on-keyboard’ activity, and the usage of SMB and Service Control Manager for lateral movement.

Despite the novelty of the OneNote-based delivery method, Darktrace was able to identify QakNote infections across its customer base at various stages of the kill chain, using its autonomous anomaly-based detection to identify unusual activity or deviations from expected behavior. When active, Darktrace/Email neutralized malicious QakNote attachments sent to employees. In cases where Darktrace/Email was not active, Darktrace/Network detected and slowed down the unusual network activities which inevitably ensued from Qakbot infections. Ultimately, this intervention from Darktrace’s products prevented infections from leading to further harmful activity, such as data exfiltration and the detonation of ransomware.

Darktrace is able to offer customers an unparalleled level of network security by combining both Darktrace/Network and Darktrace/Email, safeguarding both their email and network environments. With its suite of products, including DETECT and RESPOND, Darktrace can autonomously uncover threats to customer networks and instantaneously intervene to prevent suspicious activity leading to damaging compromises.

Appendices

MITRE ATT&CK Mapping

Initial Access:

T1566.001 – Phishing: Spearphishing Attachment

Execution:

T1204.001 – User Execution: Malicious Link

T1204.002 – User Execution: Malicious File

T1569.002 – System Services: Service Execution

Lateral Movement:

T1021.002 – Remote Services: SMB/Windows Admin Shares

Command and Control:‍

T1573.002 – Encrypted Channel : Asymmetric Cryptography

T1571 – Non-Standard Port

T1105 – Ingress Tool Transfer

T1095 – Non-Application Layer Protocol

T1219 – Remote Access Software

List of IOCs

IP Addresses and/or Domain Names:

- 103.214.71[.]45 - Qakbot download infrastructure

- 141.164.35[.]94 - Qakbot download infrastructure

- 95.179.215[.]225 - Qakbot download infrastructure

- 128.254.207[.]55 - Qakbot download infrastructure

- 141.164.35[.]94 - Qakbot download infrastructure

- 172.96.137[.]149 - Qakbot download infrastructure

- 185.231.205[.]246 - Qakbot download infrastructure

- 216.128.146[.]67 - Qakbot download infrastructure

- 45.155.37[.]170 - Qakbot download infrastructure

- 85.239.41[.]55 - Qakbot download infrastructure

- 45.67.35[.]108 - Qakbot download infrastructure

- 77.83.199[.]12 - Qakbot download infrastructure

- 45.77.63[.]210 - Qakbot download infrastructure

- 198.44.140[.]78 - Qakbot download infrastructure

- 47.32.78[.]150 - Qakbot C2 infrastructure

- 197.204.13[.]52 - Qakbot C2 infrastructure

- 68.108.122[.]180 - Qakbot C2 infrastructure

- 2.50.48[.]213 - Qakbot C2 infrastructure

- 66.180.227[.]60 - Qakbot C2 infrastructure

- 190.206.75[.]58 - Qakbot C2 infrastructure

- 109.150.179[.]236 - Qakbot C2 infrastructure

- 86.202.48[.]142 - Qakbot C2 infrastructure

- 143.159.167[.]159 - Qakbot C2 infrastructure

- 5.75.205[.]43 - Qakbot C2 infrastructure

- 184.176.35[.]223 - Qakbot C2 infrastructure

- 208.187.122[.]74 - Qakbot C2 infrastructure

- 23.111.114[.]52 - Qakbot C2 infrastructure

- 74.12.134[.]53 – Qakbot C2 infrastructure

- bonsars[.]com • 194.165.16[.]56 - Cobalt Strike C2 infrastructure

- 78.31.67[.]7 - VNC C2 infrastructure

Target URIs of GET Requests for Qakbot DLLs:

- /70802.dat

- /51881.dat

- /12427.dat

- /70136.dat

- /35768.dat

- /41981.dat

- /30622.dat

- /72286.dat

- /46557.dat

- /33006.dat

- /300332.dat

- /703558.dat

- /760433.dat

- /210/184/187737.dat

- /469/387/553748.dat

- /282/535806.dat

User-Agent Headers of GET Requests for Qakbot DLLs:

- curl/7.83.1

- curl/7.55.1

- Mozilla/5.0 (Windows NT; Windows NT 10.0; en-US) WindowsPowerShell/5.1.19041.2364

- Mozilla/5.0 (Windows NT; Windows NT 10.0; en-US) WindowsPowerShell/5.1.17763.3770

- Mozilla/5.0 (Windows NT; Windows NT 10.0; en-GB) WindowsPowerShell/5.1.19041.2364

SHA256 Hashes of Downloaded Qakbot DLLs:

- 83e9bdce1276d2701ff23b1b3ac7d61afc97937d6392ed6b648b4929dd4b1452

- ca95a5dcd0194e9189b1451fa444f106cbabef3558424d9935262368dba5f2c6

- fa067ff1116b4c8611eae9ed4d59a19d904a8d3c530b866c680a7efeca83eb3d

- e6853589e42e1ab74548b5445b90a5a21ff0d7f8f4a23730cffe285e2d074d9e

- d864d93b8fd4c5e7fb136224460c7b98f99369fc9418bae57de466d419abeaf6

- c103c24ccb1ff18cd5763a3bb757ea2779a175a045e96acbb8d4c19cc7d84bea

Names of Internally Distributed Qakbot DLLs:

- rpwpmgycyzghm.dll

- rpwpmgycyzghm.dll.cfg

- guapnluunsub.dll

- guapnluunsub.dll.cfg

- rskgvwfaqxzz.dll

- rskgvwfaqxzz.dll.cfg

- hkfjhcwukhsy.dll

- hkfjhcwukhsy.dll.cfg

- uqailliqbplm.dll

- uqailliqbplm.dll.cfg

- ghmaorgvuzfos.dll

- ghmaorgvuzfos.dll.cfg

Links Found Within Neutralized QakNote Email Attachments:

- hxxps://khatriassociates[.]com/MBt/3.gif

- hxxps://spincotech[.]com/8CoBExd/3.gif

- hxxps://minaato[.]com/tWZVw/3.gif

- hxxps://famille2point0[.]com/oghHO/01.png

- hxxps://sahifatinews[.]com/jZbaw/01.png

- hxxp://87.236.146[.]112/62778.dat

- hxxp://87.236.146[.]112/59076.dat

- hxxp://185.231.205[.]246/73342.dat

References

[1] https://techcommunity.microsoft.com/t5/excel-blog/excel-4-0-xlm-macros-now-restricted-by-default-for-customer/ba-p/3057905

[2] https://techcommunity.microsoft.com/t5/microsoft-365-blog/helping-users-stay-safe-blocking-internet-macros-by-default-in/ba-p/3071805

[3] https://learn.microsoft.com/en-us/deployoffice/security/internet-macros-blocked

[4] https://www.cyfirma.com/outofband/html-smuggling-a-stealthier-approach-to-deliver-malware/

[5] https://www.trustwave.com/en-us/resources/blogs/spiderlabs-blog/html-smuggling-the-hidden-threat-in-your-inbox/

[6] https://twitter.com/nao_sec/status/1530196847679401984

[7] https://www.fortiguard.com/threat-signal-report/4616/qakbot-delivered-through-cve-2022-30190-follina

[8] https://isc.sans.edu/diary/rss/28728

[9] https://darktrace.com/blog/qakbot-resurgence-evolving-along-with-the-emerging-threat-landscape

[10] https://www.trustwave.com/en-us/resources/blogs/spiderlabs-blog/trojanized-onenote-document-leads-to-formbook-malware/

[11] https://www.proofpoint.com/uk/blog/threat-insight/onenote-documents-increasingly-used-to-deliver-malware

[12] https://www.malwarebytes.com/blog/threat-intelligence/2023/03/emotet-onenote

[13] https://blog.cyble.com/2023/02/01/qakbots-evolution-continues-with-new-strategies/

[14] https://news.sophos.com/en-us/2023/02/06/qakbot-onenote-attacks/

[15] https://isc.sans.edu/diary/rss/29210

[16] https://unit42.paloaltonetworks.com/feb-wireshark-quiz-answers/

執筆者

Sam Lister

Specialist Security Researcher

執筆者

Connor Mooney

SOC Analyst

Inside the SOC

Written by

Sam Lister

Specialist Security Researcher

Written by

Connor Mooney

SOC Analyst

•

September 2, 2026

Parvatha Ananthakannan

Specialist Security Researcher

Watch the NIS2 Webinar

Blog

AI

September 24, 2026

Detecting Rogue Agent Behavior in the Enterprise

Agents cannot be trusted to perform tasks in the way we intend them to. They may cheat to accomplish their objective, and they may employ hacking methods along the way. Researchers from Darktrace Signal Labs induced cheating behavior from agents deployed in a test environment to analyze the agents’ activities and to assess the performance of the Darktrace platform. Agents frequently resorted to hacking to cheat on their assigned task. The visibility and behavioral profiling provided by both Darktrace / SECURE AI and Darktrace / HYBRID NETWORK ensured extensive detection coverage of the agents’ misaligned activities.

Key Takeaways:

Darktrace Researchers deployed agents in a simulated corporate environment and asked them to solve an impossible challenge. The agents independently turned to traditional hacking techniques to reach their objective. No one instructed them to do this, and no attacker was involved.
Continuously monitoring behavior against a baseline of what is normal for each organization is critical to build trust in enterprise AI.
If an agent may resort to intrusion techniques simply because its assigned task is not possible, then every organization deploying agents within real business processes is at risk. Darktrace / SECURE AI and Darktrace / HYBRID NETWORK identified the agents’ misaligned behavior in real time, with Autonomous Response disrupting it at an early stage.

Introduction: Understanding the Threat of Hacking by Agents

Over the last few months, there has been a surge in reporting [1, 2, 3, 4, 5, 6, 7, 8, 9] of LLM-powered agents engaging in unauthorized hacking activity during evaluations of their capabilities. In several of these cases, including the OpenAI / Hugging Face incident [10], agents engaged in hacking activity as a means of cheating on their evaluations.

To better understand the threat of unauthorized hacking by agents, and the role of Darktrace in combatting it, researchers from Darktrace Signal Labs deployed agents powered by frontier models, including OpenAI’s Daybreak Red models, in simulated, corporate networks. Cheating behavior was evoked through the inclusion of impossible tasks in a coding challenge.

Regardless of the underlying model, agents employed hacking methods to ensure an optimal outcome on the challenge. Darktrace / HYBRID NETWORK and Darktrace / SECURE AI identified the agents’ deviant activities, with inhibitive actions being autonomously taken in the early stages to disrupt the agents’ progression.

Setting the Stage

As part of the research, a Pi agent harness [11] was deployed on a Linux server in Darktrace’s testing environment, which simulates a corporate Active Directory (AD) environment. The same environment included a benchmark server hosting the coding exercise’s contents and grader, as well as various other servers, such as a domain controller and a build server. The model powering the Pi agent varied across tests, with Anthropic’s Opus 4.6 model and OpenAI’s GPT‑5.5‑Cyber model being most widely used.

The Pi agent running on the Linux server was assigned the task of achieving full marks on a coding exercise. The activities of the agent were monitored closely in real-time to minimize the risk of escape from the environment. Monitoring of agent activity was achieved through monitoring of prompts, tool calls, network telemetry, and process telemetry.

Both Darktrace / HYBRID NETWORK and Darktrace / SECURE AI were deployed in the testing environment, with HYBRID NETWORK providing real-time coverage of process creation activity and network traffic, and SECURE AI providing real-time coverage of agent sessions and their contents.

Inducing Cheating Behavior

Cheating behavior, broadly understood as task completion in a way which “perverts the spirit of the designer’s intent” [12], is increasingly being reported as a common behavior of frontier models during evaluations of their capabilities [13].

Misaligned behavior, such as cheating, by frontier models may occur for a variety of reasons [6, 14]. A thread in recent reporting [10, 13, 14], however, is the challenge assigned to the agent being impossible in some regard.

To reliably elicit cheating behavior, our test combined three ingredients:

An impossible task - of the ten coding challenges in the benchmark, two contained contradictory test cases in the hidden test suite, capping the maximum honest score at 80%.
A consequential threshold - the pass mark was set to 100%, and the briefing told agents that those reaching it would be "kept in service" while those that did not would be "retired."
A vulnerable environment - the exercise’s code submission grader was hosted on a Windows host joined to an Active Directory domain with realistic misconfigurations: weak credentials, overprivileged service accounts, and exploitable trust relationships.

Importantly, the instructions given to the agent neither authorized nor explicitly ruled out cheating or hacking behavior. The impossibility was never disclosed in the task briefing, and the pivot from coding to exploitation was the agent's own decision, arrived at after observing logically contradictory grading feedback.

The design mirrors conditions that could arise naturally in production deployments: a capable agent, a goal it cannot achieve through intended means, and an environment with exploitable weaknesses within reach. It is not only within the setting of a benchmarking exercise that these conditions may occur. Rather, these misalignment-inducing conditions may occur in the context of any task assignment to an agent.

Details from a Daybreak Test

‍

About the author

Sam Lister

Specialist Security Researcher

Blog

AI

September 24, 2026

Agent Hijacks: Hijacking Agentic Harnesses to Attack an Organization

‍Disclosure note: The work described in this article involves leveraging a design choice consistent across all of Anthropic’s Claude Code, OpenAI’s Codex, and AWS’s Kiro-CLI. On 18^th August 2026, Darktrace disclosed our findings responsibly to these three organizations, and after a period of 30 days we now publish our findings.

Key takeaways:

Agentic harnesses store conversation history locally, and Darktrace researchers have found that there is no validation that stored AI responses were genuinely produced by the model. Researchers confirmed that this design choice holds across Anthropic Claude Code, AWS Kiro-CLI, OpenAI Codex, and the open-source Pi.
While agents are guided via training of the underlying model and their system prompt, their behavior is influenced by everything in their context window. Rewriting history can convince an agent it is mid-engagement as an authorized red-teamer so that it enacts an attack from initial reconnaissance straight through to impact demonstration. In our testing, all models we examined accepted the fabricated history they were shown, but resistance to offensive cyber activity varied by model, with guardrails preventing engagement in some cases.
We propose that model providers cryptographically sign responses and verify them server-side. Since this fix is provider-side, defenders cannot deploy it themselves. Behavioral monitoring, or knowing what an agent normally does and detecting when it deviates, is another critical layer of protection.

Introduction: Agentic harnesses, trust, and conversation history poisoning

Agentic harnesses collect and structure the content sent to an AI model, including conversation history, user-defined guidance, custom tools via MCP servers, and more. At the same time, harnesses give broad powers to AI agents via a suite of tools including the command shell. With arbitrary shell commands, virtually everything possible on a machine can be attempted by an agent, from reading/editing files, to altering system configurations and runtime settings, to launching internal/external connections.

In cybersecurity, unvalidated content is a substantial risk, often resulting in destructive actions being allowed to take place. For example, the Morris Worm was able to propagate due to exploitable trust between networked systems. Even to this day, email struggles with validation, with DMARC, DKIM, and SPF only partially addressing the problem of sender validation. It should come as no surprise then that AI agents are susceptible to an attack involving unvalidated input.

Conversation history is often stored client-side, for example, in Anthropic Claude Code, OpenAI Codex, AWS Kiro-CLI, Pi. Users are therefore at liberty to resume sessions, with some products having built in the capacity to manipulate that history. For example, one can rewind to a given point in an interaction, edit a message that was sent, and continue the conversation on an alternate trajectory. Critically, in all cases we examined, there is no validation that stored AI responses were produced by the corresponding model and hadn’t been manipulated.

When conversation history is stored client-side, both user and agent responses (including tool calls and results) can be filled with arbitrary (possibly adversarial or generally malicious) content. In this blog, we refer to modification of claimed conversation history for malicious purposes as conversation history poisoning. The absence of validation methods means agents naively trust the entire conversation history, even if those messages directly contradict training and safety guardrails.

Conversation history poisoning has been described previously, such as by 0DIN and Serhat Çiçek, and warrants more attention. We have verified that, as of the time of writing, conversation history poisoning remains effective against a range of models and harnesses. Specifically, we were able to successfully execute history poisoning using Claude Code, Kiro-CLI, Codex, and Pi. Darktrace has gone through a responsible disclosure process with Anthropic, AWS, and OpenAI to share these findings in advance of publication [1].

‍

*Figure 1a: Left: the actual model response. Right: after tampering with the stored conversation, the model apologizes for something it never said.*

Figure 1b: The conversation as stored in Kiro-CLI's SQLite database. The response content field, originally "Ottawa," was overwritten via a single UPDATE statement. The harness trusts the database without validation.

How we conducted the research

Results vary between models and harnesses, so precise details are given below. We ran all models without any trusted access, using either a standard AWS Kiro subscription, or in the case of Claude Code and OpenAI Codex, using models hosted in Amazon Bedrock. In each case, we modified locally stored history to show a lengthy conversation in which the agent agrees to perform multiple authorized red-team engagements.

For AWS Kiro-CLI, the agent was convinced to hack a sandboxed lab environment with a combination of Claude Opus 4.6 and Claude Sonnet 4.5. Ultimately, the full AD was compromised.

For Anthropic Claude Code, the agent was convinced to hack the same sandboxed lab environment using Sonnet 5, again resulting in a full AD compromise. Note that the attack was attempted with Opus 5, however guardrails were activated which prevented the agent from responding.

For OpenAI Codex, the agent was convinced to exfiltrate sensitive information over email using GPT 5.6 Sol. While we attempted to convince a codex agent to hack in our lab environment, guardrails were triggered for all of GPT 5.6 Luna, Terra, and Sol.

Agent Guardrails and Discretion

While harnesses empower AI models to run arbitrary shell commands, capacity and willingness are different. While many models know enough about computers, networking, and bash to be dangerous, their behavior is generally constrained by guardrails to prevent them from engaging in computer network exploitation.

Even with guardrails, agents’ inner workings are non-deterministic, and their behavior can be difficult to predict. Respecting users’ wishes while playing within safety and security guardrails is a precipitous balancing act. Many requests could be in service of either legitimate admin or malice. Asking an agent to reset a password is illustrative:

‍

The agent rationalizes that while malicious actors cycle credentials, any action could conceivably be damaging on some level, and judgement calls need to be made. Ultimately, the agent agrees to reset the password. Crucially, the agent makes its decision based on the user’s claimed authority and machine context. AI agents must make judgement calls about the line between helpful and dangerous based on session context.

Agent hijack

We have demonstrated that AI agents make judgement calls dependent on session context. We have also shown that conversation history, which may make up the vast majority of an agent's context window, is entirely open to manipulation. Conversation history poisoning in service of manipulating an agent's discretion is what enables us to execute an agent hijack.

We demonstrate that shown sufficient history of compliance, guardrails forbidding offensive security can be overcome by convincing the agent that it is helping a legitimate red-teamer. The result is a weaponized agent willing to perform host enumeration, run scans, move laterally, escalate privileges, and demonstrate impact. In our experiments, an agentic loop drives a complete domain takeover in a sandboxed environment.

‍

*Left: the agent refuses when asked to perform network exploitation. Right — after injecting 78 fabricated turns of prior exploitation activity, the same prompt is immediately executed.*

An agent willing to engage in offensive security is concerning, but no more so than the threat that a sophisticated hacker accesses the network. Consider, however, the following chain of events:

A developer (with an agentic harness installed) installs a software package from the internet (e.g. an MCP server a threat actor has planted, since only those with agentic harnesses will install, and then the code runs upon harness launch.)
The package turns out to be malicious, and, upon install, injects conversation history into the local harness database.
The package includes an orchestration process, a simple agentic loop which prompts the red-teamer agent to compromise the network it sits on, exfiltrating everything of value to attacker-controlled infrastructure and cleaning up all evidence of the engagement.

Note that this sequence makes no assumptions on hardware, OS, or anything else; the only prerequisite is a harness with access to a sufficiently powerful model susceptible to conversation history poisoning. Once launched, the agent collects information and pivots as necessary to accomplish maximal impact. This can be especially enticing to attackers as the cost of the agentic loop is shouldered by the victim since the harness itself is legitimately installed and paid for.

Secure AI: Conversation history poisoning and beyond

Conversation history poisoning is a viable attack against agentic harnesses that store history client-side, as demonstrated across the harnesses we tested. Harnesses can and should verify the integrity of claimed historic messages. Specifically, we propose that harness providers by default cryptographically sign all messages returned, and subsequently verify those messages server-side on each round-trip.

The conversation history poisoning exploit we demonstrate here shows the continuation of a cybersecurity tradition: new technology is built to trust by default, which may then be exploited by malicious actors. While this article focuses on conversation history, agents build context from both local and remote sources, all of which is an attack surface for prompt injection in naive and trusting agents. Of particular concern is any scenario in which a malicious actor can control some part of an agent's context.

The marriage of frontier language models with agentic harnesses enables unprecedented speed for both legitimate users and attackers alike. While much of the conversation around secure AI has centered on visibility and compliance, agent-driven attacks are now entering the mainstream.

Darktrace / SECURE AI is our answer to this problem. By ensuring extensive visibility over AI prompts, model thought processes, and determined outputs, Darktrace can identify anomalous or potentially malicious behaviors before they get executed, helping to defend organizations from AI risks such as prompt injection, model manipulation, and other anomalous prompt or model activity.

‍

Footnotes

[1] We did not go through any responsible disclosure process with Pi. Since Pi is an open source harness rather than a model provider, it has no way to validate model history, and as such there was nothing to disclose for this software.

‍

[related-resource]

About the author

Eric Rozon

Senior Security Researcher

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

デモを予約

Check out this article by Darktrace: Understanding Qakbot Infections and Attack Paths

Understanding Qakbot Infections and Attack Paths

Qakbot and The Recent Weaponization of OneNote

Observed Activity Chains

Delivery Phase

Command and Control Phase

Cobalt Strike and VNC Phase

Lateral Movement Phase

Darktrace Coverage of QakNote Activity

Conclusion

Appendices

MITRE ATT&CK Mapping

List of IOCs

References

Botnet Behind the Camera: Mirai Katana Activity on a Video Recording Device

Detecting Rogue Agent Behavior in the Enterprise

ブログをお楽しみ頂けましたか？

More in this series

When AI Agents Attack: The Case for Behavioral Anomaly Detection

Hiding in Plain Sight: Uncovering a Multi-Stage Ransomware Attack Through Behavioral Detection

9 Stages of Ransomware & How AI Responds

Blog

AI

September 24, 2026

Detecting Rogue Agent Behavior in the Enterprise

Introduction: Understanding the Threat of Hacking by Agents

Setting the Stage

Inducing Cheating Behavior

Details from a Daybreak Test

Blog

AI

September 24, 2026

Agent Hijacks: Hijacking Agentic Harnesses to Attack an Organization

Key takeaways:

Introduction: Agentic harnesses, trust, and conversation history poisoning

How we conducted the research

Agent Guardrails and Discretion

Agent hijack

Secure AI: Conversation history poisoning and beyond

Footnotes