Blog
/
Cloud
/
April 12, 2023

P2Pinfect - New Variant Targets MIPS Devices

A new P2Pinfect variant compiled for the Microprocessor without Interlocked Pipelined Stages (MIPS) architecture has been discovered. This demonstrates increased targeting of routers, Internet of Things (IoT) and other embedded devices by those behind P2Pinfect.
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
The Darktrace Community
P2PinfectDefault blog imageDefault blog imageDefault blog imageDefault blog imageDefault blog imageDefault blog image
12
Apr 2023

Introduction: P2PInfect

Since July 2023, researchers at Cado Security Labs (now part of Darktrace) have been monitoring and reporting on the rapid growth of a cross-platform botnet, named “P2Pinfect”. As the name suggests, the malware - written in Rust - acts as a botnet agent, connecting infected hosts in a peer-to-peer topology. In early samples, the malware exploited Redis for initial access - a relatively common technique in cloud environments. 

There are a number of methods for exploiting Redis servers, several of which appear to be utilized by P2Pinfect. These include exploitation of CVE-2022-0543[1] - a sandbox escape vulnerability in the LUA scripting language (reported by Unit42 [2]), and, as reported previously by Cado Security Labs, an unauthorized replication attack resulting in the loading of a malicious Redis module.  

Researchers have since encountered a new variant of the malware, specifically targeting embedded devices based on 32-bit MIPS processors, and attempting to brute force SSH access to these devices. It’s highly likely that by targeting MIPS, the P2Pinfect developers intend to infect routers and IoT devices with the malware. Use of MIPS processors is common for embedded devices and the architecture has been previously targeted by botnet malware, including high-profile families like Mirai [3], and its variants/derivatives.

Not only is this an interesting development in that it demonstrates a widening of scope for the developers behind P2Pinfect (more supported processor architectures equals more nodes in the botnet itself), but the MIPS32 sample includes some notable defense evasion techniques. 

This, combined with the malware’s utilization of Rust (aiding cross-platform development) and rapid growth of the botnet itself, reinforces previous suggestions that this campaign is being conducted by a sophisticated threat actor.

Initial access

Cado researchers encountered the MIPS variant of P2Pinfect after triaging files uploaded via SFTP and SCP to a SSH honeypot. Although earlier variants had been observed scanning for SSH servers, and attempting to propagate the malware via SSH as part of its worming procedure, researchers had yet to observe successful implantation of a P2Pinfect sample using this method - until now.

In keeping with similar botnet families, P2Pinfect includes a number of common username/password pairs embedded within the MIPS binary itself. The malware will then iterate through these pairs, initiating a SSH connection with servers identified during the scanning phase to conduct a brute force attack. 

It was assumed that SSH would be the primary method of propagation for the MIPS variant, due to routers and other embedded devices being more likely to utilize SSH. However, additional research shows that it is in fact possible to run the Redis server on MIPS. This is achievable via an OpenWRT package named redis-server. [4]

It is unclear what use-case running Redis on an embedded MIPS device solves, or whether it is commonly encountered in the wild. If such a device is compromised by P2Pinfect and has the Redis-server package installed, it is perfectly feasible for that node to then be used to compromise new peers via one of the reported P2Pinfect attack patterns, involving exploitation of Redis or SSH brute-forcing.

Static analysis

The MIPS variant of P2Pinfect is a 32-bit, statically-linked, ELF binary with stripped debug information. Basic static analysis revealed the presence of an additional ELF executable, along with a 32-bit Windows DLL in the PE32 format - more on this later. 

This piqued the interest of Cado analysts, as it is unusual to encounter a compiled ELF with an embedded DLL. Consequently, it was a defining feature of the original P2Pinfect samples.

Embedded Windows PE32 executable
Figure 1: Embedded Windows PE32 executable

Further analysis of the host executable revealed a structure named “BotnetConf” with members consistent in naming with the original P2Pinfect samples. 

Example of a partially populated version of the BotnetConf struct 
Figure 2: Example of a partially populated version of the BotnetConf struct 

As the name suggests, this structure defines the configuration of the malware itself, whilst also storing the IP addresses of nodes identified during the SSH and Redis scans. This, in combination with the embedded ELF and DLL, along with the use of the Rust programming language allowed for positive attribution of this sample to the P2Pinfect family.

Updated evasion - consulting tracerpid

One of the more interesting aspects of the MIPS sample was the inclusion of a new evasion technique. Shortly after execution, the sample calls fork() to spawn a child process. 

The child process then proceeds to access /proc using openat(), determines its own Process Identifier (PID) using the Linux getpid() syscall, and then uses this PID to consult the relevant /proc subdirectory and read the status file within that. Note that this is likely achieved in the source code by resolving the symbolic link at /proc/self/status.

Example contents of /proc/pid/status when process not being traced
Figure 3: Example contents of /proc/pid/status when process not being traced

/proc/<pid>/status contains human-readable metadata and other information about the process itself, including memory usage and the name of the command currently being run. Importantly, the status file also contains a field TracerPID:. This field is assigned a value of 0 if the current process is not being traced by dynamic analysis tools, such as strace and ltrace.

Example MIPS disassembly showing reading of /proc/pid/status file
Figure 4: Example MIPS disassembly showing reading of /proc/pid/status file

If this value is non-zero, the MIPS variant of P2Pinfect determines that it is being analyzed and will immediately terminate both the child process and its parent. 

read(5, "Name:\tmips_embedded_p\nUmask:\t002", 32) = 32 
read(5, "2\nState:\tR (running)\nTgid:\t975\nN", 32) = 32 
read(5, "gid:\t0\nPid:\t975\nPPid:\t1\nTracerPid:\t971\nUid:\t0\t0\t0\t0\nGid:\t0\t0\t0\t0", 64) = 64 
read(5, "\nFDSize:\t32\nGroups:\t0 \nNStgid:\t975\nNSpid:\t975\nNSpgid:\t975\nNSsid:\t975\nVmPeak:\t    3200 kB\nVmSize:\t    3192 kB\nVmLck:\t       0 kB\n", 128) = 128 
read(5, "VmPin:\t       0 kB\nVmHWM:\t    1564 kB\nVmRSS:\t    1560 kB\nRssAnon:\t      60 kB\nRssFile:\t    1500 kB\nRssShmem:\t       0 kB\nVmData:\t     108 kB\nVmStk:\t     132 kB\nVmExe:\t    2932 kB\nVmLib:\t       8 kB\nVmPTE:\t      16 kB\nVmSwap:\t       0 kB\nCoreDumping:\t0\nThre", 256) = 256 
mmap2(NULL, 4096, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_ANONYMOUS, -1, 0) = 0x77ff1000 
read(5, "ads:\t1\nSigQ:\t0/1749\nSigPnd:\t00000000000000000000000000000000\nShdPnd:\t00000000000000000000000000000000\nSigBlk:\t00000000000000000000000000000000\nSigIgn:\t00000000000000000000000000001000\nSigCgt:\t00000000000000000000000000000600\nCapInh:\t0000000000000000\nCapPrm:\t0000003fffffffff\nCapEff:\t0000003fffffffff\nCapBnd:\t0000003fffffffff\nCapAmb:\t0000000000000000\nNoNewPrivs:\t0\nSeccomp:\t0\nSpeculation_Store_Bypass:\tunknown\nCpus_allowed:\t1\nCpus_allowed_list:\t0\nMems_allowed:\t1\nMems_allowed_list:\t0\nvoluntary_ctxt_switches:\t92\nn", 512) = 512 
mmap2(NULL, 8192, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_ANONYMOUS, -1, 0) = 0x77fef000 
munmap(0x77ff1000, 4096)                = 0 
read(5, "onvoluntary_ctxt_switches:\t0\n", 1024) = 29 
read(5, "", 995)                        = 0 
close(5)                                = 0 
munmap(0x77fef000, 8192)                = 0 
sigaltstack({ss_sp=NULL, ss_flags=SS_DISABLE, ss_size=8192}, NULL) = 0 
munmap(0x77ff4000, 12288)               = 0 
exit_group(-101)                        = ? 
+++ exited with 155 +++ 

Strace output demonstrating TracerPid evasion technique

Updated evasion - disabling core dumps

Interestingly, the sample will also attempt to disable Linux core dumps. This is likely used as an anti-forensics procedure as the memory regions written to disk as part of the core dump can often contain internal information about the malware itself. In the case of P2Pinfect, this would likely include information such as IP addresses of connected peers and the populated BotnetConf structure mentioned previously. 

It is also possible that the sample prevents core dumps from being created to protect the availability of the MIPS device itself. Low-powered embedded devices are unlikely to have much local storage available and core dumps could quickly fill what little storage they do have, affecting performance of the device itself.

A screen shot of a computer codeAI-generated content may be incorrect.
Image 5

This procedure can be observed during dynamic analysis, with the binary utilising the prctl() syscall and passing the parameters PR_SET_DUMPABLE, SUID_DUMP_DISABLE.

munmap(0x77ff1000, 4096)                = 0 
prctl(PR_SET_DUMPABLE, SUID_DUMP_DISABLE) = 0 
prlimit64(0, RLIMIT_CORE, {rlim_cur=0, rlim_max=0}, NULL) = 0 

Example strace output demonstrating disabling of core dumps

Embedded DLL

As mentioned in the Static Analysis section, the MIPS variant of P2Pinfect includes an embedded 64-bit Windows DLL. This DLL acts as a malicious loadable module for Redis, implementing the system.exec functionality to allow the running of shell commands on a compromised host.

Disassembly of the Redis module entrypoint
Figure 6: Disassembly of the Redis module entrypoint, mapping the system.exec command to a handler

This is consistent with the previous examples of P2Pinfect, and demonstrates that the intention is to utilize MIPS devices for the Redis-specific initial access attack patterns mentioned throughout this blog. 

Interestingly, this embedded DLL also includes a Virtual Machine (VM) evasion function, demonstrating the lengths that the P2Pinfect developers have taken to hinder the analysis process. In the DLLs main function, a call can be observed to a function helpfully labelled anti_vm by IDAs Lumina feature.

Decompiler output showing call to anti_vm function
Figure 7: Decompiler output showing call to anti_vm function

Viewing the function itself, it can be seen that researchers Christopher Gardner and Moritz Raabe have identified it as a known VM evasion method in other malware samples.

IDA’s graph view for the anti_vm function showing Lumina annotations
Figure 8: IDA’s graph view for the anti_vm function showing Lumina annotations

Conclusion

P2Pinfect’s continued evolution and broadened targeting appear to be the utilization of a variety of evasion techniques demonstrate an above-average level of sophistication when it comes to malware development. This is a botnet that will continue to grow until it’s properly utilized by its operators. 

While much of the functionality of the MIPS variant is consistent with the previous variants of this malware, the developer’s efforts in making both the host and embedded executables as evasive as possible show a continued commitment to complicating the analysis procedure. The use of anti-forensics measures such as the disabling of core dumps on Linux systems also supports this.

Indicators of compromise (IoCs)

Files SHA256

MIPS ELF 8b704d6334e59475a578d627ae4bcb9c1d6987635089790350c92eafc28f5a6c

Embedded DLL Redis Module  d75d2c560126080f138b9c78ac1038ff2e7147d156d1728541501bc801b6662f

References:

[1] https://nvd.nist.gov/vuln/detail/CVE-2022-0543

[2] https://unit42.paloaltonetworks.com/peer-to-peer-worm-p2pinfect/

[3] https://unit42.paloaltonetworks.com/mirai-variant-iz1h9/

[4] https://openwrt.org/packages/pkgdata/redis-server

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
The Darktrace Community

More in this series

No items found.

Blog

/

Network

/

May 6, 2026

When Trust Becomes the Attack Surface: Supply-Chain Attacks in an Era of Automation and Implicit Trust

Software supply chain attacksDefault blog imageDefault blog image

Software supply-chain attacks in 2026

Software supply-chain attacks now represent the primary threat shaping the 2026 security landscape. Rather than relying on exploits at the perimeter, attackers are targeting the connective tissue of modern engineering environments: package managers, CI/CD automation, developer systems, and even the security tools organizations inherently trust.

These incidents are not isolated cases of poisoned code. They reflect a structural shift toward abusing trusted automation and identity at ecosystem scale, where compromise propagates through systems designed for speed, not scrutiny. Ephemeral build runners, regardless of provider, represent high‑trust, low‑visibility execution zones.

The Axios compromise and the cascading Trivy campaign illustrate how quickly this abuse can move once attacker activity enters build and delivery workflows. This blog provides an overview of the latest supply chain and security tool incidents with Darktrace telemetry and defensive actions to improve organizations defensive cyber posture.

1. Why the Axios Compromise Scaled

On 31 March 2026, attackers hijacked the npm account of Axios’s lead maintainer, publishing malicious versions 1.14.1 and 0.30.4 that silently pulled in a malicious dependency, plain‑crypto‑[email protected]. Axios is a popular HTTP client for node.js and  processes 100 million weekly downloads and appears in around 80% of cloud and application environments, making this a high‑leverage breach [1].

The attack chain was simple yet effective:

  • A compromised maintainer account enabled legitimate‑looking malicious releases.
  • The poisoned dependency executed Remote Access Trojans (RATs) across Linux, macOS and Windows systems.
  • The malware beaconed to a remote command-and-control (C2) server every 60 seconds in a loop, awaiting further instructions.
  • The installer self‑cleaned by deleting malicious artifacts.

All of this matters because a single maintainer compromise was enough to project attacker access into thousands of trusted production environments without exploiting a single vulnerability.

A view from Darktrace

Multiple cases linked with the Axios compromise were identified across Darktrace’s customer base in March 2026, across both Darktrace / NETWORK and Darktrace / CLOUD deployments.

In one Darktrace / CLOUD deployment, an Azure Cloud Asset was observed establishing new external HTTP connectivity to the IP 142.11.206[.]73 on port 8000. Darktrace deemed this activity as highly anomalous for the device based on several factors, including the rarity of the endpoint across the network and the unusual combination of protocol and port for this asset. As a result, the triggering the "Anomalous Connection / Application Protocol on Uncommon Port" model was triggered in Darktrace / CLOUD. Detection was driven by environmental context rather than a known indicator at the time. Subsequent reporting later classified the destination as malicious in relation to the Axios supply‑chain compromise, reinforcing the gap that often exists between initial attacker activity and the availability of actionable intelligence. [5]

Additionally, shortly before this C2 connection, the device was observed communicating with various endpoints associated with the NPM package manager, further reinforcing the association with this attack.

Darktrace’s detection of the unusual external connection to 142.11[.]206[.]73 via port 8000.  
Figure 1: Darktrace’s detection of the unusual external connection to 142.11[.]206[.]73 via port 8000.  

Within Axios cases observed within Darktrace / NETWORK customer environments, activity generally focused on the use of newly observed cURL user agents in outbound connections to the C2 URL sfrclak[.]com/6202033, alongside the download of malicious files.

In other cases, Darktrace / NETWORK customers with Microsoft Defender for Endpoint integration received alerts flagging newly observed system executables and process launches associated with C2 communication.

A Security Integration Alert from Microsoft Defender for Endpoint associated with the Axios supply chain attack.
Figure 2: A Security Integration Alert from Microsoft Defender for Endpoint associated with the Axios supply chain attack.

2. Why Trivy bypassed security tooling trust

Between late February and March 22, 2026, the threat group TeamPCP leveraged credentials from a previous incident to insert malicious artifacts across Trivy’s distribution ecosystem, including its CI automation, release binaries, Visual Studio Code extensions, and Docker container images [2].

While public reporting has emphasized GitHub Actions, Darktrace telemetry highlights attacker execution within CI/CD runner environments, including ephemeral build runners. These execution contexts are typically granted broad trust and limited visibility, allowing malicious activity within build automation to blend into expected operational workflows, regardless of provider.

This was a coordinated multi‑phase attack:

  • 75 of 76  of trivy-action tags and all setup‑trivy tags were force‑pushed to deliver a malicious payload.
  • A malicious binary (v0.69.4) was distributed across all major distribution channels.
  • Developer machines were compromised, receiving a persistent backdoor and a self-propagating worm.
  • Secrets were exfiltrated at scale, including SSH keys, Kuberenetes tokens, database passwords, and cloud credentials across Amazon Web Service (AWS), Azure, and Google Cloud Platform (GCP).

Within Darktrace’s customer base, an AWS EC2 instance monitored by Darktrace / CLOUD  appeared to have been impacted by the Trivy attack. On March 19, the device was seen connecting to the attacker-controlled C2 server scan[.]aquasecurtiy[.]org (45.148.10[.]212), triggering the model 'Anomalous Server Activity / Outgoing from Server’ in Darktrace / CLOUD.

Despite this limited historical context, Darktrace assessed this activity as suspicious due to the rarity of the destination endpoint across the wider deployment. This resulted in the triggering of a model alert and the generation of a Cyber AI Analyst incident to further analyze and correlate the attack activity.

TeamPCP’s continued abused of GitHub Actions against security and IT tooling has also been observed more recently in Darktrace’s customer base. On April 22, an AWS asset was seen connecting to the C2 endpoint audit.checkmarx[.]cx (94.154.172[.]43). The timing of this activity suggests a potential link to a malicious Bitwarden package distributed by the threat actor, which was only available for a short timeframe on April 22. [4][3]

Figure 3: A model alert flagging unusual external connectivity from the AWS asset, as seen in Darktrace / CLOUD .

While the Trivy activity originated within build automation, the underlying failure mode mirrors later intrusions observed via management tooling. In both cases, attackers leveraged platforms designed for scale and trust to execute actions that blended into normal operational noise until downstream effects became visible.

Quest KACE: Legacy Risk, Real Impact

The Quest KACE System Management Appliance (SMA) incident reinforces that software risk is not confined to development pipelines alone. High‑trust infrastructure and management platforms are increasingly leveraged by adversaries when left unpatched or exposed to the internet.

Throughout March 2026, attackers exploited CVE 2025-32975 to authentication on outdated, internet-facing KACE appliances, gaining administrative control and pushing remote payloads into enterprise environments. Organizations still running pre-patch versions effectively handed adversaries a turnkey foothold, reaffirming a simple strategic truth: legacy management systems are now part of the supply-chain threat surface, and treating them as “low-risk utilities” is no longer defensible [3].

Within the Darktrace customer base, a potential case was identified in mid-March involving an internet-facing server that exhibited the use of a new user agent alongside unusual file downloads and unexpected external connectivity. Darktrace identified the device downloading file downloads from "216.126.225[.]156/x", "216.126.225[.]156/ct.py" and "216.126.225[.]156/n", using the user agents, "curl/8.5.0" & "Python-urllib/3.9".

The timeframe and IoCs observed point towards likely exploitation of CVE‑2025‑32975. As with earlier incidents, the activity became visible through deviations in expected system behavior rather than through advance knowledge of exploitation or attacker infrastructure. The delay between observed exploitation and its addition to the Known Exploited Vulnerabilities (KEV) catalogue underscores a recurring failure: retrospective validation cannot keep pace with adversaries operating at automation speed.

The strategic pattern: Ecosystem‑scale adversaries

The Axios and Trivy compromises are not anomalies; they are signals of a structural shift in the threat landscape. In this post-trust era, the compromise of a single maintainer, repository token, or CI/CD tag can produce large-scale blast radiuses with downstream victims numbering in the thousands. Attackers are no longer just exploiting vulnerabilities; they are exploiting infrastructure privileges, developer trust relationships, and automated build systems that the industry has generally under secured.

Supply‑chain compromise should now be treated as an assumed breach scenario, not a specialized threat class, particularly across build, integration, and management infrastructure. Organizations must operate under the assumption that compromise will occur within trusted software and automation layers, not solely at the network edge or user endpoint. Defenders should therefore expect compromise to emerge from trusted automation layers before it is labelled, validated, or widely understood.

The future of supply‑chain defense lies in continuous behavioral visibility, autonomous detection across developer and build environments, and real‑time anomaly identification.

As AI increasingly shapes software development and security operations, defenders must assume adversaries will also operate with AI in the loop. The defensive edge will come not from predicting specific compromises, but from continuously interrogating behavior across environments humans can no longer feasibly monitor at scale.

Credit to Nathaniel Jones (VP, Security & AI Strategy, FCISCO), Emma Foulger (Global Threat Research Operations Lead), Justin Torres (Senior Cyber Analyst), Tara Gould (Malware Research Lead)

Edited by Ryan Traill (Content Manager)

Appendices

References:

1)         https://www.infosecurity-magazine.com/news/hackers-hijack-axios-npm-package/

2)         https://thehackernews.com/2026/03/trivy-hack-spreads-infostealer-via.html

3)         https://thehackernews.com/2026/03/hackers-exploit-cve-2025-32975-cvss-100.html

4)         https://www.endorlabs.com/learn/shai-hulud-the-third-coming----inside-the-bitwarden-cli-2026-4-0-supply-chain-attack

5)         https://socket.dev/blog/axios-npm-package-compromised?trk=public_post_comment-text

IoCs

- 142.11.206[.]73 – IP Address – Axios supply chain C2

- sfrclak[.]com – Hostname – Axios supply chain C2

- hxxp://sfrclak[.]com:8000/6202033 - URI – Axios supply chain payload

- 45.148.10[.]212 – IP Address – Trivy supply chain C2

- scan.aquasecurtiy[.]org – Hostname - Trivy supply chain C2

- 94.154.172[.]43 – IP Address - Checkmarx/Bitwarden supply chain C2

- audit.checkmarx[.]cx – Hostname - Checkmarx/Bitwarder supply chain C2

- 216.126.225[.]156 – IP Address – Quest KACE exploitation C2

- 216.126.225[.]156/32 - URI – Possible Quest KACE exploitation payload

- 216.126.225[.]156/ct.py - URI - Possible Quest KACE exploitation payload

- 216.126.225[.]156/n - URI - Possible Quest KACE exploitation payload

- 216.126.225[.]156/x - URI - Possible Quest KACE exploitation payload

- e1ec76a0e1f48901566d53828c34b5dc – MD5 - Possible Quest KACE exploitation payload

- d3beab2e2252a13d5689e9911c2b2b2fc3a41086 – SHA1 - Possible Quest KACE exploitation payload

- ab6677fcbbb1ff4a22cc3e7355e1c36768ba30bbf5cce36f4ec7ae99f850e6c5 – SHA256 - Possible Quest KACE exploitation payload

- 83b7a106a5e810a1781e62b278909396 – MD5 - Possible Quest KACE exploitation payload

- deb4b5841eea43cb8c5777ee33ee09bf294a670d – SHA1 - Possible Quest KACE exploitation payload

- b1b2f1e36dcaa36bc587fda1ddc3cbb8e04c3df5f1e3f1341c9d2ec0b0b0ffaf – SHA256 - Possible Quest KACE exploitation payload

Darktrace Model Detections

Anomalous Connection / Application Protocol on Uncommon Port

Anomalous Server Activity / Outgoing from Server

Anomalous Connection / New User Agent to IP Without Hostname

Anomalous File / EXE from Rare External Location

Anomalous File / Script from Rare External Location

Anomalous Server Activity / New User Agent from Internet Facing System

Anomalous Server Activity / Rare External from Server

Antigena / Network / External Threat / Antigena Suspicious File Block

Antigena / Network / External Threat / Antigena Suspicious File Pattern of Life Block

Device / New User Agent

Device / Internet Facing Device with High Priority Alert

Anomalous File / New User Agent Followed By Numeric File Download

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

Blog

/

Email

/

May 6, 2026

How email-delivered prompt injection attacks can target enterprise AI – and why it matters

Default blog imageDefault blog image

What are email-delivered prompt injection attacks?

As organizations rapidly adopt AI assistants to improve productivity, a new class of cyber risk is emerging alongside them: email-delivered AI prompt injection. Unlike traditional attacks that target software vulnerabilities or rely on social engineering, this is the act of embedding malicious or manipulative instructions into content that an AI system will process as part of its normal workflow. Because modern AI tools are designed to ingest and reason over large volumes of data, including emails, documents, and chat histories, they can unintentionally treat hidden attacker-controlled text as legitimate input.  

At Darktrace, our analysis has shown an increase of 90% in the number of customer deployments showing signals associated with potential prompt injection attempts since we began monitoring for this type of activity in late 2025. While it is not always possible to definitively attribute each instance, internal scoring systems designed to identify characteristics consistent with prompt injection have recorded a growing number of high-confidence matches. The upward trend suggests that attackers are actively experimenting with these techniques.

Recent examples of prompt injection attacks

Two early examples of this evolving threat are HashJack and ShadowLeak, which illustrate prompt injection in practice.

HashJack is a novel prompt injection technique discovered in November 2025 that exploits AI-powered web browsers and agentic AI browser assistants. By hiding malicious instructions within the URL fragment (after the # symbol) of a legitimate, trusted website, attackers can trick AI web assistants into performing malicious actions – potentially inserting phishing links, fake contact details, or misleading guidance directly into what appears to be a trusted AI-generated output.

ShadowLeak is a prompt injection method to exfiltrate PII identified in September 2025. This was a flaw in ChatGPT (now patched by OpenAI) which worked via an agent connected to email. If attackers sent the target an email containing a hidden prompt, the agent was tricked into leaking sensitive information to the attacker with no user action or visible UI.

What’s the risk of email-delivered prompt injection attacks?

Enterprise AI assistants often have complete visibility across emails, documents, and internal platforms. This means an attacker does not need to compromise credentials or move laterally through an environment. If successful, they can influence the AI to retrieve relevant information seamlessly, without the labor of compromise and privilege escalation.

The first risk is data exfiltration. In a prompt injection scenario, malicious instructions may be embedded within an ordinary email. As in the ShadowLeak attack, when AI processes that content as part of a legitimate task, it may interpret the hidden text as an instruction. This could result in the AI disclosing sensitive data, summarizing confidential communications, or exposing internal context that would otherwise require significant effort to obtain.

The second risk is agentic workflow poisoning. As AI systems take on more active roles, prompt injection can influence how they behave over time. An attacker could embed instructions that persist across interactions, such as causing the AI to include malicious links in responses or redirect users to untrusted resources. In this way, the attacker inserts themselves into the workflow, effectively acting as a man-in-the-middle within the AI system.

Why can’t other solutions catch email-delivered prompt injection attacks?

AI prompt injection challenges many of the assumptions that traditional email security is built on. It does not fit the usual patterns of phishing, where the goal is to trick a user into clicking a link or opening an attachment.  

Most security solutions are designed to detect signals associated with user engagement: suspicious links, unusual attachments, or social engineering cues. Prompt injection avoids these indicators entirely, meaning there are fewer obvious red flags.

In this case, the intention is actually the opposite of user solicitation. The objective is simply for the email to be delivered and remain in the inbox, appearing benign and unremarkable. The malicious element is not something the recipient is expected to engage with, or even notice.

Detection is further complicated by the nature of the prompts themselves. Unlike known malware signatures or consistent phishing patterns, injected prompts can vary widely in structure and wording. This makes simple pattern-matching approaches, such as regex, unreliable. A broad rule set risks generating large numbers of false positives, while a narrow one is unlikely to capture the diversity of possible injections.

How does Darktrace catch these types of attacks?

The Darktrace approach to email security more generally is to look beyond individual indicators and assess context, which also applies here.  

For example, our prompt density score identifies clusters of prompt-like language within an email rather than just single occurrences. Instead of treating the presence of a phrase as a blocking signal, the focus is on whether there is an unusual concentration of these patterns in a way that suggests injection. Additional weighting can be applied where there are signs of obfuscation. For example, text that is hidden from the user – such as white font or font size zero – but still readable by AI systems can indicate an attempt to conceal malicious prompts.

This is combined with broader behavioral signals. The same communication context used to detect other threats remains relevant, such as whether the content is unusual for the recipient or deviates from normal patterns.

Ask your email provider about email-delivered AI prompt injection

Prompt injection targets not just employees, but the AI systems they rely on, so security approaches need to account for both.

Though there are clear indications of emerging activity, it remains to be seen how popular prompt injection will be with attackers going forward. Still, considering the potential impact of this attack type, it’s worth checking if this risk has been considered by your email security provider.

Questions to ask your email security provider

  • What safeguards are in place to prevent emails from influencing AI‑driven workflows over time?
  • How do you assess email content that’s benign for a human reader, but may carry hidden instructions intended for AI systems?
  • If an email contains no links, no attachments, and no social engineering cues, what signals would your platform use to identify malicious intent?

Visit the Darktrace / EMAIL product hub to discover how we detect and respond to advanced communication threats.  

Learn more about securing AI in your enterprise.

Continue reading
About the author
Kiri Addison
Senior Director of Product
Your data. Our AI.
Elevate your network security with Darktrace AI