Blog
/
Network
/
July 26, 2022

Identifying PrivateLoader Network Threats

Learn how Darktrace identifies network-based indicators of compromise for the PrivateLoader malware. Gain insights into advanced threat detection.
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
Shuh Chin Goh
Default blog image
26
Jul 2022

Instead of delivering their malicious payloads themselves, threat actors can pay certain cybercriminals (known as pay-per-install (PPI) providers) to deliver their payloads for them. Since January 2022, Darktrace’s SOC has observed several cases of PPI providers delivering their clients’ payloads using a modular malware downloader known as ‘PrivateLoader’.

This blog will explore how these PPI providers installed PrivateLoader onto systems and outline the steps which the infected PrivateLoader bots took to install further malicious payloads. The details provided here are intended to provide insight into the operations of PrivateLoader and to assist security teams in identifying PrivateLoader bots within their own networks.  

Threat Summary 

Between January and June 2022, Darktrace identified the following sequence of network behaviours within the environments of several Darktrace clients. Patterns of activity involving these steps are paradigmatic examples of PrivateLoader activity:

1. A victim’s device is redirected to a page which instructs them to download a password-protected archive file from a file storage service — typically Discord Content Delivery Network (CDN)

2. The device contacts a file storage service (typically Discord CDN) via SSL connections

3. The device either contacts Pastebin via SSL connections, makes an HTTP GET request with the URI string ‘/server.txt’ or ‘server_p.txt’ to 45.144.225[.]57, or makes an HTTP GET request with the URI string ‘/proxies.txt’ to 212.193.30[.]45

4. The device makes an HTTP GET request with the URI string ‘/base/api/statistics.php’ to either 212.193.30[.]21, 85.202.169[.]116, 2.56.56[.]126 or 2.56.59[.]42

5. The device contacts a file storage service (typically Discord CDN) via SSL connections

6. The device makes a HTTP POST request with the URI string ‘/base/api/getData.php’ to either 212.193.30[.]21, 85.202.169[.]116, 2.56.56[.]126 or 2.56.59[.]42

7. The device finally downloads malicious payloads from a variety of endpoints

The PPI Business 

Before exploring PrivateLoader in more detail, the pay-per-install (PPI) business should be contextualized. This consists of two parties:  

1. PPI clients - actors who want their malicious payloads to be installed onto a large number of target systems. PPI clients are typically entry-level threat actors who seek to widely distribute commodity malware [1]

2. PPI providers - actors who PPI clients can pay to install their malicious payloads 

As the smugglers of the cybercriminal world, PPI providers typically advertise their malware delivery services on underground web forums. In some cases, PPI services can even be accessed via Clearnet websites such as InstallBest and InstallShop [2] (Figure 1).  

Figure 1: A snapshot of the InstallBest PPI login page [2]


To utilize a PPI provider’s service, a PPI client must typically specify: 

(A)  the URLs of the payloads which they want to be installed

(B)  the number of systems onto which they want their payloads to be installed

(C)  their geographical targeting preferences. 

Payment of course, is also required. To fulfil their clients’ requests, PPI providers typically make use of downloaders - malware which instructs the devices on which it is running to download and execute further payloads. PPI providers seek to install their downloaders onto as many systems as possible. Follow-on payloads are usually determined by system information garnered and relayed back to the PPI providers’ command and control (C2) infrastructure. PPI providers may disseminate their downloaders themselves, or they may outsource the dissemination to third parties called ‘affiliates’ [3].  

Back in May 2021, Intel 471 researchers became aware of PPI providers using a novel downloader (dubbed ‘PrivateLoader’) to conduct their operations. Since Intel 471’s public disclosure of the downloader back in Feb 2022 [4], several other threat research teams, such as the Walmart Cyber Intel Team [5], Zscaler ThreatLabz [6], and Trend Micro Research [7] have all provided valuable insights into the downloader’s behaviour. 

Anatomy of a PrivateLoader Infection

The PrivateLoader downloader, which is written in C++, was originally monolithic (i.e, consisted of only one module). At some point, however, the downloader became modular (i.e, consisting of multiple modules). The modules communicate via HTTP and employ various anti-analysis methods. PrivateLoader currently consists of the following three modules [8]: 

  • The loader module: Instructs the system on which it is running to retrieve the IP address of the main C2 server and to download and execute the PrivateLoader core module
  • The core module: Instructs the system on which it is running to send system information to the main C2 server, to download and execute further malicious payloads, and to relay information regarding installed payloads back to the main C2 server
  • The service module: Instructs the system on which it is running to keep the PrivateLoader modules running

Kill Chain Deep-Dive 

The chain of activity starts with the user’s browser being redirected to a webpage which instructs them to download a password-protected archive file from a file storage service such as Discord CDN. Discord is a popular VoIP and instant messaging service, and Discord CDN is the service’s CDN infrastructure. In several cases, the webpages to which users’ browsers were redirected were hosted on ‘hero-files[.]com’ (Figure 2), ‘qd-files[.]com’, and ‘pu-file[.]com’ (Figure 3). 

Figure 2: An image of a page hosted on hero-files[.]com - an endpoint which Darktrace observed systems contacting before downloading PrivateLoader from Discord CDN
Figure 3: An image of a page hosted on pu-file[.]com- an endpoint which Darktrace observed systems contacting before downloading PrivateLoader from Discord CDN

‍
On attempting to download cracked/pirated software, users’ browsers were typically redirected to download instruction pages. In one case however, a user’s device showed signs of being infected with the malicious Chrome extension, ChromeBack [9], immediately before it contacted a webpage providing download instructions (Figure 4). This may suggest that cracked software downloads are not the only cause of users’ browsers being redirected to these download instruction pages (Figure 5). 

Figure 4: The event log for this device (taken from the Darktrace Threat Visualiser interface) shows that the device contacted endpoints associated with ChromeBack ('freychang[.]fun') prior to visiting a page ('qd-file[.]com') which instructed the device’s user to download an archive file from Discord CDN
 Figure 5: An image of the website 'crackright[.]com'- a provider of cracked software. Systems which attempted to download software from this website were subsequently led to pages providing instructions to download a password-protected archive from Discord CDN

‍
After users’ devices were redirected to pages instructing them to download a password-protected archive, they subsequently contacted cdn.discordapp[.]com over SSL. The archive files which users downloaded over these SSL connections likely contained the PrivateLoader loader module. Immediately after contacting the file storage endpoint, users’ devices were observed either contacting Pastebin over SSL, making an HTTP GET request with the URI string ‘/server.txt’ or ‘server_p.txt’ to 45.144.225[.]57, or making an HTTP GET request with the URI string ‘/proxies.txt’ to 212.193.30[.]45 (Figure 6).

Distinctive user-agent strings such as those containing question marks (e.g. ‘????ll’) and strings referencing outdated Chrome browser versions were consistently seen in these HTTP requests. The following chrome agent was repeatedly observed: ‘Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/74.0.3729.169 Safari/537.36’.

In some cases, devices also displayed signs of infection with other strains of malware such as the RedLine infostealer and the BeamWinHTTP malware downloader. This may suggest that the password-protected archives embedded several payloads.

Figure 6: This figure, obtained from Darktrace's Advanced Search interface, represents the post-infection behaviour displayed by a PrivateLoader bot. After visiting hero-files[.]com and downloading the PrivateLoader loader module from Discord CDN, the device can be seen making HTTP GET requests for ‘/proxies.txt’ and ‘/server.txt’ and contacting pastebin[.]com

It seems that PrivateLoader bots contact Pastebin, 45.144.225[.]57, and 212.193.30[.]45 in order to retrieve the IP address of PrivateLoader’s main C2 server - the server which provides PrivateLoader bots with payload URLs. This technique used by the operators of PrivateLoader closely mirrors the well-known espionage tactic known as ‘dead drop’.

The dead drop is a method of espionage tradecraft in which an individual leaves a physical object such as papers, cash, or weapons in an agreed hiding spot so that the intended recipient can retrieve the object later on without having to come in to contact with the source. When threat actors host information about core C2 infrastructure on intermediary endpoints, the hosted information is analogously called a ‘Dead Drop Resolver’ or ‘DDR’. Example URLs of DDRs used by PrivateLoader:

  • https://pastebin[.]com/...
  • http://212.193.30[.]45/proxies.txt
  • http://45.144.225[.]57/server.txt
  • http://45.144.255[.]57/server_p.txt

The ‘proxies.txt’ DDR hosted on 212.193.40[.]45 contains a list of 132 IP address / port pairs. The 119th line of this list includes a scrambled version of the IP address of PrivateLoader’s main C2 server (Figures 7 & 8). Prior to June, it seems that the main C2 IP address was ‘212.193.30[.]21’, however, the IP address appears to have recently changed to ‘85.202.169[.]116’. In a limited set of cases, Darktrace also observed PrivateLoader bots retrieving payload URLs from 2.56.56[.]126 and 2.56.59[.]42 (rather than from 212.193.30[.]21 or 85.202.169[.]116). These IP addresses may be hardcoded secondary C2 address which PrivateLoader bots use in cases where they are unable to retrieve the primary C2 address from Pastebin, 212.193.30[.]45 or 45.144.255[.]57 [10]. 

Figure 7: Before June, the 119th entry of the ‘proxies.txt’ file lists '30.212.21.193' -  a scrambling of the ‘212.193.30[.]21’ main C2 IP address
Figure 8: Since June, the 119th entry of the ‘proxies.txt’ file lists '169.85.116.202' - a scrambling of the '85.202.169[.]116' main C2 IP address

Once PrivateLoader bots had retrieved C2 information from either Pastebin, 45.144.225[.]57, or 212.193.30[.]45, they went on to make HTTP GET requests for ‘/base/api/statistics.php’ to either 212.193.30[.]21, 85.202.169[.]116, 2.56.56[.]126, or 2.56.59[.]42 (Figure 9). The server responded to these requests with an XOR encrypted string. The strings were encrypted using a 1-byte key [11], such as 0001101 (Figure 10). Decrypting the string revealed a URL for a BMP file hosted on Discord CDN, such as ‘hxxps://cdn.discordapp[.]com/attachments/978284851323088960/986671030670078012/PL_Client.bmp’. These encrypted URLs appear to be file download paths for the PrivateLoader core module. 

Figure 9: HTTP response from server to an HTTP GET request for '/base/api/statistics.php'
Figure 10: XOR decrypting the string with the one-byte key, 00011101, outputs a URL in CyberChef

After PrivateLoader bots retrieved the 'cdn.discordapp[.]com’ URL from 212.193.30[.]21, 85.202.169[.]116, 2.56.56[.]126, or 2.56.59[.]42, they immediately contacted Discord CDN via SSL connections in order to obtain the PrivateLoader core module. Execution of this module resulted in the bots making HTTP POST requests (with the URI string ‘/base/api/getData.php’) to the main C2 address (Figures 11 & 12). Both the data which the PrivateLoader bots sent over these HTTP POST requests and the data returned via the C2 server’s HTTP responses were heavily encrypted using a combination of password-based key derivation, base64 encoding, AES encryption, and HMAC validation [12]. 

Figure 11: The above image, taken from Darktrace's Advanced Search interface, shows a PrivateLoader bot carrying out the following steps: contact ‘hero-files[.]com’ --> contact ‘cdn.discordapp[.]com’ --> retrieve ‘/proxies.txt’ from 212.193.30[.]45 --> retrieve ‘/base/api/statistics.php’ from 212.193.30[.]21 --> contact ‘cdn.discordapp[.]com --> make HTTP POST request with the URI ‘base/api/getData.php’ to 212.193.30[.]21
Figure 12: A PCAP of the data sent via the HTTP POST (in red), and the data returned by the C2 endpoint (in blue)

These ‘/base/api/getData.php’ POST requests contain a command, a campaign name and a JSON object. The response may either contain a simple status message (such as “success”) or a JSON object containing URLs of payloads. After making these HTTP connections, PrivateLoader bots were observed downloading and executing large volumes of payloads (Figure 13), ranging from crypto-miners to infostealers (such as Mars stealer), and even to other malware downloaders (such as SmokeLoader). In some cases, bots were also seen downloading files with ‘.bmp’ extensions, such as ‘Service.bmp’, ‘Cube_WW14.bmp’, and ‘NiceProcessX64.bmp’, from 45.144.225[.]57 - the same DDR endpoint from which PrivateLoader bots retrieved main C2 information. These ‘.bmp’ payloads are likely related to the PrivateLoader service module [13]. Certain bots made follow-up HTTP POST requests (with the URI string ‘/service/communication.php’) to either 212.193.30[.]21 or 85.202.169[.]116, indicating the presence of the PrivateLoader service module, which has the purpose of establishing persistence on the device (Figure 14). 

Figure 13: The above image, taken from Darktrace's Advanced Search interface, outlines the plethora of malware payloads downloaded by a PrivateLoader bot after it made an HTTP POST request to the ‘/base/api/getData.php’ endpoint. The PrivateLoader service module is highlighted in red
Figure 14: The event log for a PrivateLoader bot, obtained from the Threat Visualiser interface, shows a device making HTTP POST requests to ‘/service/communication.php’ and connecting to the NanoPool mining pool, indicating successful execution of downloaded payloads

In several observed cases, PrivateLoader bots downloaded another malware downloader called ‘SmokeLoader’ (payloads named ‘toolspab2.exe’ and ‘toolspab3.exe’) from “Privacy Tools” endpoints [14], such as ‘privacy-tools-for-you-802[.]com’ and ‘privacy-tools-for-you-783[.]com’. These “Privacy Tools” domains are likely impersonation attempts of the legitimate ‘privacytools[.]io’ website - a website run by volunteers who advocate for data privacy [15]. 

After downloading and executing malicious payloads, PrivateLoader bots were typically seen contacting crypto-mining pools, such as NanoPool, and making HTTP POST requests to external hosts associated with SmokeLoader, such as hosts named ‘host-data-coin-11[.]com’ and ‘file-coin-host-12[.]com’ [16]. In one case, a PrivateLoader bot went on to exfiltrate data over HTTP to an external host named ‘cheapf[.]link’, which was registered on the 14th March 2022 [17]. The name of the file which the PrivateLoader bot used to exfiltrate data was ‘NOP8QIMGV3W47Y.zip’, indicating information stealing activities by Mars Stealer (Figure 15) [18]. By saving the HTTP stream as raw data and utilizing a hex editor to remove the HTTP header portions, the hex data of the ZIP file was obtained. Saving the hex data using a ‘.zip’ extension and extracting the contents, a file directory consisting of system information and Chrome and Edge browsers’ Autofill data in cleartext .txt file format could be seen (Figure 16).

Figure 15: A PCAP of a PrivateLoader bot’s HTTP POST request to cheapf[.]link, with data sent by the bot appearing to include Chrome and Edge autofill data, as well as system information
Figure 16: File directory structure and files of the ZIP archive 

When left unattended, PrivateLoader bots continued to contact C2 infrastructure in order to relay details of executed payloads and to retrieve URLs of further payloads. 

Figure 17: Timeline of the attack

Darktrace Coverage 

Most of the incidents surveyed for this article belonged to prospective customers who were trialling Darktrace with RESPOND in passive mode, and thus without the ability for autonomous intervention. However in all observed cases, Darktrace DETECT was able to provide visibility into the actions taken by PrivateLoader bots. In one case, despite the infected bot being disconnected from the client’s network, Darktrace was still able to provide visibility into the device’s network behaviour due to the client’s usage of Darktrace/Endpoint. 

If a system within an organization’s network becomes infected with PrivateLoader, it will display a range of anomalous network behaviours before it downloads and executes malicious payloads. For example, it will contact Pastebin or make HTTP requests with new and unusual user-agent strings to rare external endpoints. These network behaviours will generate some of the following alerts on the Darktrace UI:

  • Compliance / Pastebin 
  • Device / New User Agent and New IP
  • Device / New User Agent
  • Device / Three or More New User Agents
  • Anomalous Connection / New User Agent to IP Without Hostname
  • Anomalous Connection / POST to PHP on New External Host
  • Anomalous Connection / Posting HTTP to IP Without Hostname

Once the infected host obtains URLs for malware payloads from a C2 endpoint, it will likely start to download and execute large volumes of malicious files. These file downloads will usually cause Darktrace to generate some of the following alerts:

  • Anomalous File / EXE from Rare External Location
  • Anomalous File / Numeric Exe Download
  • Anomalous File / Masqueraded File Transfer
  • Anomalous File / Multiple EXE from Rare External Locations
  • Device / Initial Breach Chain Compromise

If RESPOND is deployed in active mode, Darktrace will be able to autonomously block the download of additional malware payloads onto the target machine and the subsequent beaconing or crypto-mining activities through network inhibitors such as ‘Block matching connections’, ‘Enforce pattern of life’ and ‘Block all outgoing traffic’. The ‘Enforce pattern of life’ action results in a device only being able to make connections and data transfers which Darktrace considers normal for that device. The ‘Block all outgoing traffic’ action will cause all traffic originating from the device to be blocked. If the customer has Darktrace’s Proactive Threat Notification (PTN) service, then a breach of an Enhanced Monitoring model such as ‘Device / Initial Breach Chain Compromise’ will result in a Darktrace SOC analyst proactively notifying the customer of the suspicious activity. Below is a list of Darktrace RESPOND (Antigena) models which would be expected to breach due to PrivateLoader activity. Such models can seriously hamper attempts made by PrivateLoader bots to download malicious payloads. 

  • Antigena / Network / External Threat / Antigena Suspicious File Block
  • Antigena / Network / Significant Anomaly / Antigena Controlled and Model Breach
  • Antigena / Network / External Threat / Antigena File then New Outbound Block
  • Antigena / Network / Significant Anomaly / Antigena Significant Anomaly from Client Block 
  • Antigena / Network / Significant Anomaly / Antigena Breaches Over Time Block

In one observed case, the infected bot began to download malicious payloads within one minute of becoming infected with PrivateLoader. Since RESPOND was correctly configured, it was able to immediately intervene by autonomously enforcing the device’s pattern of life for 2 hours and blocking all of the device’s outgoing traffic for 10 minutes (Figure 17). When malware moves at such a fast pace, the availability of autonomous response technology, which can respond immediately to detected threats, is key for the prevention of further damage.  

Figure 18: The event log for a Darktrace RESPOND (Antigena) model breach shows Darktrace RESPOND performing inhibitive actions once the PrivateLoader bot begins to download payloads

Conclusion

By investigating PrivateLoader infections over the past couple of months, Darktrace has observed PrivateLoader operators making changes to the downloader’s main C2 IP address and to the user-agent strings which the downloader uses in its C2 communications. It is relatively easy for the operators of PrivateLoader to change these superficial network-based features of the malware in order to evade detection [19]. However, once a system becomes infected with PrivateLoader, it will inevitably start to display anomalous patterns of network behaviour characteristic of the Tactics, Techniques and Procedures (TTPs) discussed in this blog.

Throughout 2022, Darktrace observed overlapping patterns of network activity within the environments of several customers, which reveal the archetypal steps of a PrivateLoader infection. Despite the changes made to PrivateLoader’s network-based features, Darktrace’s Self-Learning AI was able to continually identify infected bots, detecting every stage of an infection without relying on known indicators of compromise. When configured, RESPOND was able to immediately respond to such infections, preventing further advancement in the cyber kill chain and ultimately preventing the delivery of floods of payloads onto infected devices.

IoCs

‍

MITRE ATT&CK Techniques Observed

References

[1], [8],[13] https://www.youtube.com/watch?v=Ldp7eESQotM  

[2] https://news.sophos.com/en-us/2021/09/01/fake-pirated-software-sites-serve-up-malware-droppers-as-a-service/

[3] https://www.researchgate.net/publication/228873118_Measuring_Pay-per Install_The_Commoditization_of_Malware_Distribution 

[4], [15] https://intel471.com/blog/privateloader-malware

[5] https://medium.com/walmartglobaltech/privateloader-to-anubis-loader-55d066a2653e 

[6], [10],[11], [12] https://www.zscaler.com/blogs/security-research/peeking-privateloader 

[7] https://www.trendmicro.com/en_us/research/22/e/netdooka-framework-distributed-via-privateloader-ppi.html

[9] https://www.gosecure.net/blog/2022/02/10/malicious-chrome-browser-extension-exposed-chromeback-leverages-silent-extension-loading/

[14] https://www.proofpoint.com/us/blog/threat-insight/malware-masquerades-privacy-tool 

[16] https://asec.ahnlab.com/en/30513/ 

[17]https://twitter.com/0xrb/status/1515956690642161669

[18] https://isc.sans.edu/forums/diary/Arkei+Variants+From+Vidar+to+Mars+Stealer/28468

[19] http://detect-respond.blogspot.com/2013/03/the-pyramid-of-pain.html

‍

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
Shuh Chin Goh

More in this series

No items found.

Blog

/

/

September 24, 2026

Detecting Rogue Agent Behavior in the Enterprise

Default blog imageDefault blog image

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:

  1. 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%.
  2. 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."
  3. 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

‍

‍

Continue reading
About the author
Sam Lister
Specialist Security Researcher

Blog

/

/

September 24, 2026

Agent Hijacks: Hijacking Agentic Harnesses to Attack an Organization

Default blog imageDefault blog image

‍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 18th 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:

  1. 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.)
  2. The package turns out to be malicious, and, upon install, injects conversation history into the local harness database.
  3. 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]

Continue reading
About the author
Eric Rozon
Senior Security Researcher
Your data. Our AI.
Elevate your network security with Darktrace AI