Blog
/
Email
/
May 1, 2026

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

Prompt injection is a newly emerging threat, with only a handful of confirmed victims so far – targeting how AI systems use data rather than exploiting traditional software vulnerabilities. As agentic AI becomes embedded across enterprise environments, attackers may attempt to manipulate these systems through hidden instructions in everyday email content.
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
Kiri Addison
Senior Director of Product
Default blog image
01
May 2026

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.

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
Kiri Addison
Senior Director of Product

More in this series

No items found.

Blog

/

/

November 5, 2026

From One Alert to a Wider Campaign: Uncovering Novel Threats Through Behavioral Analysis

Default blog imageDefault blog image

Key Findings

  • Darktrace’s Threat Research team is exploring a new method to identify malicious campaigns by correlating similar anomalous activity across its customer fleet.
  • The method uses unsupervised clustering to identify shared behavioral patterns across multiple customer deployments without relying on prior indicators of compromise (IoCs).
  • The approach identified shared infrastructure across dozens of customers, with activity linked to the likely POPA botnet.
  • A second cluster identified anomalous activity associated with Monero cryptocurrency mining.

‍

How does Darktrace conduct threat research?

The Darktrace Threat Research team aims to identify malicious campaign activity affecting multiple customers. By their nature, these campaigns will lead to shared behavioral patterns across multiple victims.

In this piece, we describe a method for identifying malicious campaign activity that uses this simple observation as a guiding principle. We begin by outlining a few key facts about the modern threat landscape and why behavioral detection is necessary to meet the challenges it presents. We then explore how fleet-wide behavioral detections can be leveraged using the principle above to systematically identify multi-deployment campaigns impacting the Darktrace customer fleet. Finally, we demonstrate that this research is not purely theoretical, using real examples of campaign activity detected through this novel approach.

Shifting to a behavioral-focused approach

The modern era of cyber-defense necessitates detection capabilities beyond traditional signature-based methods. Technologies such as antivirus signatures, hash-based malware identification, and IP blocklists have historically been successful in defending against known threats. However, they are easily circumvented by attackers and inherently limited in defending against the novel attack techniques that today's threat landscape is producing.

Recent breakthroughs in AI represent a transformational risk multiplier that cannot be ignored. AI systems do not merely execute existing attacks at machine speed; there are signs they can also synthesize novel tradecraft [1], increasing the likelihood that defenders will encounter attack patterns that have never previously been observed. Meanwhile vulnerability disclosures continue to accelerate. More than 48,000 CVEs were published in 2025, and published statistics to date indicate that 2026 is already on track to significantly exceed that figure [2]. While multiple factors may contribute to vulnerability discovery, AI is seen as a key factor in this rapid growth. Organizations should expect a growing proportion of malicious activity to fall outside the scope of detections that rely solely on prior knowledge of known threats.

By design, Darktrace’s behavioral and anomaly-focused detection capabilities are well-suited to this testing landscape. With a focus on detecting unknown and unexpected activity that deviates from the baseline of “normal” behavior, detection of CVE exploitation without prior knowledge of the vulnerability and before public disclosure, for example, is achievable [3][4].

Darktrace’s diverse customer base benefits from a further layer of protection: working with only deidentified alert metadata, our threat researchers hunt proactively across the fleet. Beyond the level of the individual customer, there is a second, larger vantage to be leveraged.

Darktrace data provides unique insight into the unexpected behavioral patterns that occur across its customer fleet. What can we see if we look at this fleet-wide telemetry and hunt for systematic behavioral campaigns that impact our customer base?

Correlating across the customer base

At the deployment level, Darktrace focuses on whether activity is unusual in the context of an organization's normal behavior, but as we zoom out and look at these alerts at the fleet level, the problem fundamentally changes . We are no longer seeking anomalies in largely normal data; we are seeking systematic patterns within data already known to be anomalous. An isolated behavior anomaly may be notable for a single customer; the same fingerprint across dozens is a signal of a structured campaign. Hunting for such campaigns and sharing our findings are key areas of focus for Darktrace’s Threat Research team.  In this spirit, Darktrace is  actively exploring a a new research approach that leverages deidentified alert metadata to increase campaign detection ability.

This approach is based on the premise that an expert human analyst presented with a set of alerts will identify likely related activity by assessing whether the metadata fields indicate a relationship. For example, two connections to a rare endpoint that exhibits naming structure that is likely from a single algorithm are a strong indicator, as is the use of an unusual port carrying a given protocol (e.g. port 6000 for SSL). Darktrace’s approach to campaign detection encodes this expert human judgement process through deterministic embedding functions that obtain reproducible vector representations of alert data. Unsupervised clustering is then applied to these vector representations to obtain groups of alerts that describe semantically similar alert data across multiple customers. These clusters therefore represent similar anomalous activity that has occurred across multiple customers. In other words, these clusters are candidates for malicious campaign activity.

‍

Figure 1: Layers of AI-based analysis provided through Darktrace products and cross-customer correlation.

‍

Case study 1 – unusual connections to shared underlying infrastructure

In one recent case, the clustering produced several tight groups of alerts spanning dozens of customers. Each group centered on a single domain whose subdomains were algorithmically generated (e.g. s01699, s1736, s91). Examined side by side, these clusters shared an unmistakable behavioral shape: each involved a domain with a clear naming pattern (e.g. s016841786.net-echo[.]com, s1670.pixellog[.]io, s217.pixellog[.]io) , with connections over SSL on the unusual port 6000 to servers hosted by non-mainstream providers (e.g. OVH, Hetzner, GTHost, Leaseweb), rather than generally expected content-delivery networks.

‍

Figure 2: An example model alert triggering on unusual outbound connections and data transfer to s1670.pixellog[.]io.

‍

Closer inspection revealed that the same subdomain label appeared under different domains (e.g. s1786.net-echo[.]com and s1786.pixellog[.]io). Following this lead, Darktrace’s threat researchers pivoted across the customer base and found numerous distinct subdomain labels reused under more than one registered domain. In several cases, the same label resolved to the same IP across different domains (e.g. s1704 under both pixellog[.]io and net-echo[.]com → 148.113.222[.]244; s01684 → 15.235.67[.]95), suggesting the use of ephemeral subdomain infrastructure in an attempt to avoid signature-based defense techniques.

This example demonstrates how multi-deployment data can uncover shared infrastructure that may be difficult to identify at the individual deployment level alone. In particular, the presented method facilitated this discovery by producing coherent, tight clusters of geometrically similar vector embeddings of malicious activity from large-scale fleet data, which could then be presented to a researcher for further investigation. This  activity highlighted across the Darktrace customer base was ultimately linked to likely POPA botnet activity through use of open-source intelligence (OSINT) [5].

Case Study 2 – anomalous user agent and txt URI

A second example shows the method surfacing a single, self-contained cluster. The clustering was not given any indicators to look for; instead, it grouped devices that shared an unusual combination of features. On inspection, that shared consisted of:

-      unusual user agent (tes)

-      specific request path (/md5_GIP.txt)

-      uncommon destination port (83)

-      hostname impersonating Google (matching regular expression ^www\.google\.com\.[a-zA-Z0-9]+\.fast-dns-host\.com$)

Treating these shared features as IoCs, Darktrace’s threat researchers then pivoted to an exploration across the customer base to investigate the full breadth of the campaign. This search identified further commonalities; additional domains which shared the same IP as the domain family highlighted in the cluster, as well as evidence of cryptocurrency mining on affected devices. Additional OSINT research also aligned with these findings, linking the discovered payloads and infrastructure to Monero cryptocurrency mining [6].

Conclusion

The method outlined in this blog is under active development, with ongoing technical refinement providing opportunities to further enhance its capabilities. While the capability continues to mature, this approach to threat hunting leverages the powerful vantage point of fleetwide telemetry across a diverse customer base to help protect Darktrace customers. This method allows campaigns that at the deployment level appear as isolated alerts to be seen at their true scale and spread. It enables proactive fleet-wide analysis without prior IoCs to identify campaigns impacting Darktrace customers.

This is possible because of what the method is built on: the atomic unit is the Darktrace alert, a signal that encodes years of cyber expertise, while the multi-deployment nature of the method allows organizations to indirectly benefit from findings from other networks through shared intelligence by Darktrace’s Threat Research team.

Credit to Daniel Levy (Threat Hunting Data Scientist), Paul Jennings (Principal Cyber Analyst), Emma Foulger (Global Head, Threat Operations)

Edited by Ryan Traill (Content Manager)

References

1. https://www.aisi.gov.uk/blog/incident-report-unsanctioned-agent-behaviour-during-cyber-testing

2. https://www.cvefind.com/en/cve-statistic-date.html

3. https://www.darktrace.com/blog/security-after-signatures-operating-in-a-world-of-pre-cve-disclosure-exploitation-collapsed-trust-boundaries-and-autonomous-systems

4. https://www.darktrace.com/blog/pre-cve-threat-detection-10-examples-identifying-malicious-activity-prior-to-public-disclosure-of-a-vulnerability

5. https://github.com/synthient/public-research/blob/main/2026/06/popa/Popa.md

6. https://www.deepwatch.com/blog/pirated-adobe-photoshop-2023-installers-lead-to-monero-cryptocurrency-mining/

Continue reading
About the author
Daniel Levy
Threat Hunting Data Scientist

Blog

/

AI

/

November 2, 2026

PHP Vulnerabilities, ARTEX and the Growing Role of AI in Cyber Attacks

Default blog imageDefault blog image

The disclosure of CVE-2026-109618 and CVE-2026-109619 placed immediate pressure on organizations running PHP. Both vulnerabilities can allow unauthenticated remote code execution, including in products and appliances where PHP is embedded, and updates depend on third-party vendors [1].

Early guidance focused on patching internet-facing systems and monitoring for behavior associated with successful exploitation, including unexpected command execution, Webroot changes and outbound connections from PHP servers.

Darktrace observed widespread scanning across monitored environments involving hundreds of distinct source IP addresses that public reporting has associated with activity targeting the PHP vulnerabilities. The volume of scanning was significantly greater than the evidence available for successful exploitation. For analysts, that distinction is important. Scanning demonstrates interest in exposed infrastructure; establishing compromise requires evidence of subsequent behavior on the targeted system, including command execution, file modification or unexpected outbound communication.

PHP is not an isolated example. Imperva subsequently demonstrated remote code execution against a vulnerable Node.js build through malicious Brotli data, reproducing execution through compressed HTTP responses and configured request parsers. The same inputs were rejected by the patched runtime. Imperva linked its Node2Shell and PHP2Shell research to the increasing speed of vulnerability discovery and exploit development. [2]

While Darktrace has not observed direct evidence of CVE-2026-109618 or CVE-2026-109619 exploitation, on October 21st we observed a six-month high in detections consistent with zero-day exploitation activity, representing a 64% increase compared to historic data. This increase has not been confirmed as related to the PHP vulnerabilities. Darktrace’s behavioral approach is designed to detect anomalous activity without relying solely on rules written for a specific exploit. Comparable increases were observed around the Log4j disclosure and other globally significant vulnerability disclosures.

AI-assisted vulnerability research

The PHP disclosure arrives as AI systems have become more capable at vulnerability research [3]. OpenAI launched Daybreak in June 2026, combining GPT-5.5-Cyber, Codex Security and partnerships with security researchers and software maintainers. OpenAI says its models have been applied to discovering and generating patches for vulnerabilities in browsers, network infrastructure, FreeBSD and the Linux kernel. Darktrace joined the OpenAI Daybreak Cyber Partner Program in June to explore how OpenAI's cyber capabilities could be integrated within Darktrace products and services [4][10].

OpenAI reported in August that Codex Security Cloud had analyzed more than 30 million commits across more than 30,000 codebases. According to OpenAI, Cloudflare uses models accessed through Daybreak to investigate customer-authorized code and combines findings with live traffic and security signals to assess whether affected routes are active or receiving suspicious traffic, with proposed mitigations and patches remaining subject to customer review [5][6].

Faster discovery still leaves defenders with the work of locating affected software, testing fixes and completing remediation. The PHP case shows that gap clearly, particularly when vulnerable runtimes are embedded inside other products.

Enterprise agents as an attack surface

Darktrace Signal Labs has also examined how a trusted enterprise agent could be abused after an attacker has already obtained access to an environment. In a controlled simulation, researchers demonstrated that, in that post-compromise scenario, the Claude platform’s memory stores feature could be abused to establish a live C2 channel between an external operator and internal Claude Code agents. The external operator was able to dynamically assign network discovery, credential access, and data exfiltration tasks to the agents via this C2 channel. Detection of the agents’ activities relied on the correlation of process creation and network traffic activity from agent harness processes, alongside behavioral analysis to identify deviations from expected behavior. Darktrace has not claimed it observed this memory-store command-and-control technique in operational attacks [7].

AI-assisted attacks in Japan

Evidence from Japan indicates AI being used for vulnerability scanning and attacks outside controlled research. On 21 October, JPCERT/CC published findings from an investigation into recent unauthorized access incidents affecting organizations in Japan. JPCERT/CC states that it confirmed attackers had used AI to scan for known vulnerabilities and conduct attacks through APIs. According to JPCERT/CC, its examination of attack infrastructure found evidence of Android application analysis, API discovery and testing for weaknesses including server-side request forgery, and the infrastructure also targeted services intended for users outside Japan [8].

JPCERT/CC does not identify ARTEX by name in that publication. The named ARTEX attribution comes from separate investigations, including Microsoft's analysis of Storm-0940.

AI in offensive operations

Microsoft describes ARTEX as an open-source, multi-agent penetration-testing system capable of automating activity with limited human interaction. Microsoft reported that Storm-0940, which it assesses to be a China state-aligned actor active since 2021, used ARTEX against hundreds of organizations, and Microsoft attributed the ConnectWise supply-chain compromise [9].

Microsoft observed ARTEX accelerating scanning, reconnaissance, password spraying and OAuth application abuse. Several ARTEX-driven operations reached privileged cloud access less than three hours after initial targeting. Following the ConnectWise compromise, Storm-0940 attempted to authenticate to more than 2,250 tenants at rates of up to 2,900 attempts per minute [9].

Darktrace Threat Research found activity in its own telemetry involving infrastructure later associated with the campaign. On 13 October, Darktrace recorded failed authentication activity involving 144.172.64[.]128, one of the addresses Microsoft later identified as performing sign-ins and Microsoft Graph activity in ConnectWise customer environments. Darktrace analysis identified 65 failed authentication attempts in around 20 minutes.

Figure 1: Darktrace Advanced Search showing failed authentication events involving 144.172.64[.]128 on 13 October 2026, (customer-identifying information redacted).

The telemetry establishes repeated failed authentication from infrastructure later associated with the campaign. It does not establish ARTEX execution or successful compromise. At the time of observation, Microsoft's later attribution was unavailable. Analysts therefore had the behavior itself: repeated failed authentication from the same source over a short period, together with any related identity, application or network activity inside the environment.

Microsoft documented a wider password-spray campaign against 327 accounts across eight organizations between 10 and 14 October. It recorded more than 151,000 failed logons, including more than 111,000 on 11 October, but no successful authentication, token issuance, Graph call or directory change from that campaign. Microsoft also reported that most Storm-0940 targeting was unsuccessful [9].

Where the actor progressed beyond authentication, Microsoft observed service-principal sign-ins followed by Graph and directory operations. Activity included reconnaissance, application credential changes and attempts to modify privileges. Microsoft recommends correlating service-principal sign-ins with Graph activity, directory changes and resource access [9].

Behavior before attribution

The PHP and ARTEX cases begin with different information. PHP defenders knew which software required urgent remediation but initially had limited information about exploitation. Process execution, file modification and outbound communication could still provide evidence of compromise.

The PHP and ARTEX observations also illustrate the difference between volume and impact. Darktrace observed broad scanning around the PHP disclosure, while the ARTEX-associated infrastructure produced concentrated authentication activity. Neither observation independently demonstrates compromise and in each case examining the behavior that follows the initial interaction is the investigative clue.

The Darktrace ARTEX-related observation began with authentication behavior. Microsoft publicly attributed the source infrastructure eight days later. That attribution added context, but it did not change what had originally been recorded in the environment.

AI is now present at several points in this cycle. Daybreak applies models to vulnerability discovery and remediation. Darktrace Signal Labs demonstrated how a trusted enterprise agent could be co-opted in a controlled environment. JPCERT/CC confirmed AI-assisted vulnerability scanning and API attacks in recent incidents. Microsoft documented ARTEX being used by Storm-0940 in active operations [4][7][8][9].

Microsoft's findings also show why scale and impact must be assessed separately. One ARTEX-associated campaign produced more than 151,000 failed logons without successful access, while other operations reached privileged cloud access in less than three hours. Security teams have to distinguish high-volume attempts from activity that progresses through authentication, privilege changes and resource access.

The evidence required to assess these cases remains familiar: authentication records, process execution, application and directory changes, file activity and network communication. AI can increase the rate at which those behaviors occur. Defenders still need to establish their sequence and impact.

Credit to Emma Foulger, Sam Lister, Dylan Evans, Moishe Tajtelbaum

References

[1] NCSC-NL, php2shell vulnerabilities (CVE-2026-109618, CVE-2026-109619), 21 October 2026. https://github.com/NCSC-NL/php2shell

[2] Imperva, https://www.imperva.com/blog/node2shell-from-brotli-data-to-remote-code-execution/

[3] Cyber Security Agency of Singapore, https://www.csa.gov.sg/alerts-and-advisories/alerts/al-2026-151/

[4] OpenAI, Daybreak: Tools for securing every organization in the world, 22 June 2026. https://openai.com/index/daybreak-securing-the-world/

[5] OpenAI Developers, Scaling cyber defenders with Daybreak, 21 August 2026. https://developers.openai.com/blog/scaling-cyber-defenders-with-daybreak

[6] OpenAI, Cloudflare links vulnerability discovery to defense with OpenAI, 21 October 2026. https://openai.com/index/cloudflare-cyber/

[7] Darktrace Signal Labs, Shared Memories as a Covert C2 Channel, draft research supplied for review.

[8] JPCERT/CC, Recent AI-Assisted Cyber Attacks in Japan: Investigating Attack Infrastructure and Techniques, 21 October 2026. https://blogs.jpcert.or.jp/en/2026/10/web_api_attacks.html

[9] Microsoft Security Research and Microsoft Threat Intelligence, AI-driven, large-scale nation-state campaign using ARTEX enables supply chain compromise, 21 October 2026. https://www.microsoft.com/en-us/security/blog/2026/10/21/storm-0940-uses-artex-agentic-ai-connectwise-supply-chain-compromise/

[10] Darktrace, Advancing the Use of Frontier AI in Cybersecurity: Darktrace Joins the OpenAI Daybreak Cyber Partner Program to Explore Defensive AI Integrations, 22 June 2026. https://www.darktrace.com/blog/advancing-the-use-of-frontier-ai-in-cybersecurity-darktrace-joins-the-openai-daybreak-cyber-partner-program-to-explore-defensive-ai-integrations

Continue reading
About the author
Nathaniel Jones
SVP, Global Threat Intelligence
Your data. Our AI.
Elevate your network security with Darktrace AI