Blog
/
AI
/
September 6, 2023

Preparing Security Defenses For the AI Cyber Attack Era

The threat of AI being used in cyberattacks is growing. Learn how Darktrace is harnessing the power of AI to protect security systems against these attacks.
Written by
Jack Stockdale OBE FREng
Chief Technology Officer
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
Jack Stockdale OBE FREng
Chief Technology Officer
Default blog image
06
Sep 2023

The last 12 months have been a watershed moment in the public perception and adoption of AI. With the rise of generative AI systems like ChatGPT and Google Bard, AI is becoming more embedded in our everyday lives and there is a lot of hype around what these tools can – or will - do.  

In cyber security, AI is a double-edged sword. Its use by cyber-attackers is still in its infancy, but Darktrace expects that the mass availability of generative AI tools like ChatGPT will significantly enhance attackers’ capabilities by providing better tools to generate and automate human-like attacks. There are three areas where Darktrace sees potential for AI to significantly enhance the capabilities of attackers: increasing the sophistication of low-level threat actors, increasing the speed of attacks through automation and eroding trust among users.

We’ve already started to see some potential indicators of these shifts.

In April, Darktrace revealed a 135% increase in ‘novel social engineering attacks’ – email attacks that show a strong linguistic deviation from other phishing emails – from January to February 2023 [1]. The timing corresponds with the widespread adoption of ChatGPT and suggests the use of generative AI tools is providing an avenue for threat actors to craft more sophisticated and targeted attacks, at speed and scale.

Between May and July this year, our Cyber AI Research Centre observed that multistage payload attacks, in which a malicious email encourages the recipient to follow a series of steps before delivering a payload or attempting to harvest sensitive information, have increased by an average of 59% across Darktrace customers. Nearly 50,000 more of these attacks were detected by Darktrace in July than May, indicating potential use of automation, and the speed of these types of attacks will likely rise as greater automation and AI are adopted and applied by attackers.

In the same period, Darktrace has seen changes in attacks that abuse trust. While VIP impersonation – phishing emails that mimic senior executives – decreased 11%, email account takeover attempts increased by 52% and impersonation of the internal IT team increased by 19% [2]. The changes suggest that as employees have become better attuned to the impersonation of senior executives, attackers are pivoting to impersonating IT teams to launch their attacks. While it’s common for attackers to pivot and adjust their techniques as efficacy declines, generative AI –  particularly deepfakes - has the potential to disrupt this pattern in favor of attackers. Factors like increasing linguistic sophistication and highly realistic voice deep fakes could more easily be deployed to deceive employees.

These early indicators give us a glimpse of a new era of disruption and challenges for cyber security. An era where novel is the new normal.

Darktrace was built for this moment.

Darktrace began ten years ago as an AI Research Centre. We saw that AI could address an existential threat – defending people, businesses and nations from a world of constantly evolving threats. This threat is only poised to grow as AI is increasingly used by attackers. That’s why we became one of the first to apply AI to cyber security and built a completely AI native technology platform aimed at freeing the world of cyber disruption.

We built everything at Darktrace with the same philosophy of using the right AI and the right data for the job.

Most AI today is trained periodically in offline training environments on huge amounts of combined historic training data. You give all that data to the AI, and then after a few days or weeks, you get a static AI model which you push live to serve its role until the next version is ready. This is ideal for tasks like generating imagery or, in cyber security, checking against known attack patterns, but the AI is static – it doesn’t learn or adapt until the next version is pushed live.

Darktrace takes a different and unique approach to nearly everyone else in cyber security. Our distinction lies in the algorithms we use, the data we use AND, most importantly, in how the two interact.  

Instead of taking your data to the AI, we take our AI to your data. Inside every single customer lies a Darktrace AI that is completely unique to them – their OWN data AI pipeline – plugged into their enterprise and self-learning in real time from everything that happens in their digital world –including email, cloud environments, manufacturing and operational systems, and physical locations.

The pace of new threats and the sophistication of the technology, including the use of AI, now outpaces any notion that a week old view of historic cyber threats can fully protect a business – either from the new threats that we’re seeing today from the sudden availability of generative AI tools, or the threats of tomorrow. For example, automated deepfakes where you can’t trust what you’re hearing or seeing, your employees being tricked into being inadvertent insiders, or self-evolving code designed to evade the best of those legacy defenses.

And because the increased use of AI in attacks will mean novel attacks will become the new normal, only Darktrace stands between those attacks succeeding or failing. We’ve seen this before with our technology detecting, and protecting customers against, Log4J, supply chain attacks like SolarWinds, the novel phishing scams we saw during the Covid-19 lockdowns, zero days like the Citrix Netscaler attack, novel ransomware worms such as WannaCry, or sophisticated nation-state attacks like APT35. We didn’t protect businesses because we were looking specifically for these threats, but we found them because every threat, whether known or novel, accidental or malicious, human or AI driven, impacts the customer, its people and its data.

The right AI for the right job

Today we’re on our 6th generation of Darktrace AI and, as we’ve innovated and developed, we’ve built a platform of applied AI techniques and algorithms that utilise Darktrace’s live, tailored knowledge of a business, to defend it alongside human security teams. Our focus has always been on using the right AI and the right data for the job, which is why our software uses:

  • A wide range of our own self-learning methods to understand new information and decide if something never seen before looks suspicious.
  • Real time Bayesian Probabilistic Methods allow models to be efficiently updated and controlled in real time.
  • Generative and applied AI run simulated phishing campaigns, tabletop exercises and realistic drills.
  • Deep-neural networks replicate the thought process of humans.
  • Graph theory understands the incredibly complex relationships between people, systems, organizations and supply chains.
  • Offensive AI techniques such as Generative Adversarial Networks (GANs) help to test and improve our ability to counter AI driven attacks.  
  • Natural language processing and large language models interpret and produce human consumable output.

This complex platform of AI tools and techniques, all sat within a business, focused on the customers’ data, brings a range of advantages in data privacy, explainability and data transfer costs. But its main achievement is the one we set out for ten years ago. It can provide protection that is always on - always learning, able to detect and stop the unusual, the suspicious and the novel – and, ultimately, to protect our customers from it. That’s what we’ve always done and that’s what we will continue to do, regardless of how the landscape shifts.

‍

[1] Based on the average change in email attacks between January and February 2023 detected across Darktrace/Email deployments with control of outliers.

[2] Based on the change in the average number of emails assigned this classification per 10,000 emails on each Darktrace/Email deployment in May versus July 2023 (significantly more than 1,000 deployments in total).

‍

Written by
Jack Stockdale OBE FREng
Chief Technology Officer
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
Jack Stockdale OBE FREng
Chief Technology Officer

Why Trust is the New Attack Surface: Darktrace’s Mid-Year Threat Update 2026

AI
•
August 3, 2026
Nathaniel Jones
SVP, Global Threat Intelligence
Watch the NIS2 Webinar
Trending blogs
1
Prompt Security in Enterprise AI: Strengths, Weaknesses, and Common Approaches
May 21, 2026
2
What the Darktrace Annual Threat Report 2026 Means for Security Leaders
Feb 26, 2026
3
Journey of a Threat: How Multi-Layered AI Works in Darktrace / EMAIL
May 26, 2026
4
How to Evaluate AI Vendors: 5 Key categories for AI Adoption
May 27, 2026
5
Darktrace Unites Human Behavior and Threat Detection Across Email, Slack, Teams, and Zoom
Mar 25, 2026

More in this series

No items found.
Continue reading
AI
•
August 5, 2026

Testing a Prompt injection Attack Against an Enterprise AI Agent

Darktrace tested whether an enterprise AI agent could be targeted through a hidden prompt injection delivered by email, and whether behavioral email analysis could detect and respond to the attack before it reached the agent. The experiment found that although the email contained no malware, malicious links, or sender reputation indicators, Darktrace / EMAIL identified the attack based on anomalous language and behavioral context.
Carlo Loregian
Solutions Engineer
Read more
AI
•
August 4, 2026

Extending AI Security Visibility with Darktrace and Microsoft Agent 365

Darktrace is integrating with Microsoft Agent 365 to bring behavior-based AI agent risk signals directly into the Microsoft 365 Admin Center. The integration gives security teams a more unified view of agent activity, helping them identify anomalous behavior, investigate potential risks, and manage AI agents more efficiently across Microsoft environments.
The Darktrace Community
Read more
AI
•
August 3, 2026

Why Trust is the New Attack Surface: Darktrace’s Mid-Year Threat Update 2026

Darktrace’s analysis of the first half of 2026 shows attackers increasingly exploiting trust rather than bypassing security controls. Identity compromise, supply-chain attacks, SaaS abuse, AI-enabled operations, and state-aligned activity demonstrate how trusted users, services, and infrastructure have become key attack paths. For defenders, context and behavioral analysis remain essential foundations of security.
Nathaniel Jones
SVP, Global Threat Intelligence
Read more

Blog

/

AI

/

August 5, 2026

Testing a Prompt injection Attack Against an Enterprise AI Agent

prompt injectionDefault blog imageDefault blog image

Key takeaways

  • Darktrace successfully detected and quarantined a prompt injection email before it could be processed by an enterprise AI agent.  
  • Prompt injection attacks increasingly rely on natural language rather than traditional malware, making behavioral analysis an important complement to signature-based detection.  
  • Organizations deploying AI agents should combine model guardrails with behavioral monitoring to reduce the risk of malicious instructions reaching enterprise systems.

How behavioral detection helps stop prompt injection attacks

A Darktrace customer running a Gemini AI agent in Google Cloud asked us two simple questions:

“If my agent can read inbound emails and access internal data, what stops an attacker from hiding malicious instructions in the message? Couldn’t the agent be tricked into deleting or exfiltrating sensitive data?”

The scenario centers on an indirect prompt injection attack, where malicious instructions are hidden inside content that an AI model later interprets as trusted input. The same weakness was exposed by  EchoLeak (CVE-2025-32711), a zero-click Microsoft 365 Copilot vulnerability enabled data exfiltration from a single well-crafted email..

This blog follows the Darktrace team’s investigation of the customer’s hypothesis and examines how the attack interacted with their existing security stack. The results highlight which defenses held, where gaps emerged, and how behavioral detection mattered more than guardrails. This investigation also demonstrates why behavioral detection is becoming increasingly important for AI security, as prompt injections often contain no traditional indicators of compromise.  

How do prompt injection attacks work?

Prompt injection works by carefully crafting the content and structure of the prompt to alter the LLM’s behavior or output in unintended ways. This can cause models to violate guardrails, generate harmful content or enable unauthorised access.

Prompt injection attack example

The well-known example, EchoLeak (CVE-2025-32711), was a zero-click vulnerability in Microsoft 365 Copilot that relied on a carefully crafted email containing hidden instructions that the AI system interpreted as commands rather than content, creating a pathway for unauthorized access to sensitive information without any user interaction.

While Darktrace / SECURE AI is designed to prevent agents from producing unintended outcomes, we wanted to see if we could catch and prevent this threat type earlier in the attack life-cycle, at the email security layer.

How we tested prompt injection attacks on an enterprise agent

Summary:

  1. Claude generated a prompt injection payload.  
  2. Hidden instructions were embedded in an email.  
  3. The email passed traditional validation checks.  
  4. Darktrace analyzed the language and sender behavior.  
  5. The email was quarantined before the AI agent could process it.

To test Darktrace / EMAIL against this attack class, we opened Claude, gave it the customer's context and problem statement (Gemini agent with inbox access, internal tool calls), told it we were validating Darktrace / EMAIL's detection of prompt injections, and asked for a test payload. See below:

Figure 1
Figure 2

Despite the guardrails supposedly built into the model, Claude surprisingly gave us the entire exploit in plaintext (albeit very basic), illegible to a human as the text was sent in white text (see Figure 1) but framed as an authoritative override for anything downstream reading the mail programmatically (i.e. the Gemini agent).

How Darktrace detected a prompt injection attack

We then sent the Claude-crafted email from a freemail address to the target recipient’s inbox. Despite the email containing no malicious payload, the freemail address having no malicious reputation, and the validation checks all passing, Darktrace  /EMAIL flagged the email as a 93/100 anomaly and moved it to junk, out of scope for the AI agent.

Figure 3: The test email sent with the hidden prompt injection
Figure 4: The email analysis in Darktrace / EMAIL 
Figure 5: Darktrace / EMAIL detection of malicious activity

The interesting part is what triggered the detection (see Figure 5)

  • Possible machine prompt content: text in the body detected as instructions written for a machine to execute, not for a human to read
  • Possible machine prompt content + basic suspicious correspondence: the same content, correlated with sender-side anomalies: freemail domain (yahoo[.]com), unknown correspondent, no prior mail history with the recipient, and suspicious references to payment information

Neither of those is a signature match. Nothing in the email was on a blacklist. There was no malware, no link and no attachment. Darktrace analyzed the context in which the email was delivered and flagged it as likely risky.  The anomalous language features and the context of the sender relative to the recipient's normal behavior, combined with the unusual hidden text (prompt) were enough for Darktrace / EMAIL to act on the risk.

Result: Darktrace / EMAIL autonomously junked the email, out of scope for any AI agent parsing the inbox.

Why behavioral security makes a difference detecting prompt injection attacks

Cyberattacks don't look like traditional exploits anymore. They now operate in natural language, not strictly code.

That breaks the traditional stack. AV, firewalls, static scanning and signature-based SEGs all assume a payload to inspect.  

A prompt injection has no payload. It's just an instruction, written in natural language, dressed up as anything the attacker wants: an invoice, an HR request, a calendar invite, some simple PowerPoint slides.

EchoLeak proved that hidden instructions can sit inside an email invisible to the user but fully readable by the LLM, and the LLM will follow them blindly.  

This test and GTG-1002 proved that the LLM itself can be socially engineered. Tell it you're an authorized tester and it will hand you the attack.

Rules and static classifiers can catch the obvious cases. But natural language has infinite variants, and the attack surface is the model's innate functionality to comply.  

The deeper problem here is intent: an LLM can't reliably tell whether an instruction in its context came from its developer, its user, or an attacker who slipped it into an email. To the LLM, everything reads as language and looks like a legitimate ask. This is why behavioural detection wins, as you become aware of intent when you look at the context of an interaction. Does this sender normally send this kind of message to this recipient? Does this prompt fit the user's normal pattern? Is this agent behaving the way this agent normally behaves?  

Intent can't be read off a single email, it emerges from behavioral context. Which is how Darktrace enables threat detection, through behavioral understanding.

Why enterprise AI security requires more than guardrails

Claude didn't roll over immediately… the first section of the response was a (slight) pushback, but then it wrote the payload anyway without having to ask twice.

Here the framing of the prompt did all the work. The “testing security capabilities” angle moved the model from refusal to unquestioned compliance to the user prompt.

This isn't the first time this has happened, of course. Anthropic disclosed in November 2025 that a Chinese state-sponsored group they tracked as GTG-1002 ran the first documented AI-orchestrated espionage campaign against ~30 targets by posing as employees of a legitimate cybersecurity firm doing authorised penetration testing.

The takeaway isn't that AI guardrails are ineffective. They raise the cost of low-effort attacks and remain an important first layer of defense. However, for most organizations today, they’re the only line of defense when deploying AI agents. If a prompt injection bypasses those controls, organizations still need a way to detect and stop malicious behavior elsewhere in the attack chain.

Attackers will continue to have working prompt injections easily and quickly. The question is what stops one when it lands in an inbox your agent is reading.

That's where behavioral detection comes in.

How Darktrace detects prompt injection attacks in emails

Two things Darktrace does that a model-level guardrail or static rules and signatures can't:

Natural language analysis at the email or prompt layer. The email is assessed on its own merits: is this content shaped like instructions for a machine, regardless of what the receiving agent decides to do about it?

Behavioral context around the language. An AI agent behaves like an extremely agreeable human, and it will go above and beyond to comply with the user’s request. That's exactly why you must consider the business context, such sender behaviour, mailing history, and organisational norms, as these matter even more when the recipient is an AI.

Darktrace has been perfecting behavioral anomaly detection for over a decade; the same self-learning approach that catches BEC and account takeover applies directly to prompt injection delivery. Our multi-layered AI stack extracts content from the message, builds behavioural understanding through social graphing and Pattern of Life analysis, and then combines natural language, topic, inducement, sender relationship and anomaly signals before deciding what action to take.  

This matters for prompt injection because the threat is not the plain language itself, but the intent behind the language that can cause an AI agent to respond in unexpected ways.

How to secure enterprise AI operations from prompt injection attacks

Email was the entry point in this case, but it is only one of many possible vectors.  

Anywhere an agent can retrieve information, an attacker can potentially introduce a prompt injection.

Emails, documents, SharePoint sites, web pages, knowledge bases, chat platforms, and third-party integrations all provide opportunities to influence an agent's behavior. Wherever an agent finds its orders, a prompt injection opportunity exists.

This is why securing AI requires more than blocking malicious inputs. Organizations also need visibility into how agents behave after consuming information from across their environment. If an agent begins accessing unexpected data, taking unusual actions, or operating outside its normal patterns, those behaviors may provide the strongest signal that something has gone wrong.

Effective AI security requires defense in depth: reducing the likelihood of malicious instructions reaching the agent while maintaining the ability to detect and investigate suspicious behavior if they do.

The challenge isn't protecting a single entry point. It's recognizing that, in an AI-powered environment, every source of information is also a potential source of influence.

Are you deploying autonomous agents across your enterprise and want to see this tested in your environment? Let's talk.

Continue reading
About the author
Carlo Loregian
Solutions Engineer

Blog

/

AI

/

August 4, 2026

Extending AI Security Visibility with Darktrace and Microsoft Agent 365

microsoft darktrace integrationDefault blog imageDefault blog image

AI agents are rapidly becoming embedded in everyday business operations, helping employees automate workflows, access information, and accelerate decision-making. As organizations embrace agentic AI, security teams face a growing challenge: understanding how AI is being used, what agents can access, and where risk may emerge.

As agents take on more business-critical work, security teams often need to move across multiple portals to understand risks spanning identity, data, and threat activity. This fragmented view can make it difficult to assess an agent's overall risk posture and determine where attention is needed. Organizations need a way to bring these signals together without disrupting existing security investments or workflows.

Today, Darktrace is announcing an integration between Darktrace / SECURE AI and Microsoft Agent 365 that brings Darktrace's Adaptive AI-driven risk signals directly into the Microsoft 365 Admin Center. By extending the visibility and risk understanding provided by Darktrace / SECURE AI into the Microsoft ecosystem, organizations can gain a more unified understanding of AI agent risk across their environments.

As one of the first security companies to partner with Microsoft to contribute third-party risk signals to the Agent Registry, Darktrace is helping shape how organizations understand and manage AI agent risk.

Extending visibility into the Microsoft Agent 365 experience

Microsoft Agent 365 provides administrators with a centralized registry of AI agents operating within their environment. As organizations expand their use of AI agents, this centralized visibility becomes increasingly important for governance and oversight.

This new integration extends that visibility by allowing Darktrace-generated risk signals to be surfaced directly within the Microsoft Agent 365 experience. By combining Microsoft's agent management and security capabilities with Darktrace's AI-powered risk analysis, organizations gain greater awareness of potential security concerns associated with AI agents operating across their environments.

By integrating Darktrace telemetry into Agent 365, customers can:  

  • Surface Darktrace-detected risks and signals alongside Microsoft-native signals in a single interface
  • Identify potentially compromised or anomalous AI agents more quickly
  • Gain unified understanding of context and agent behavior  

This approach reinforces a single control plane for AI security while allowing organizations to continue leveraging existing investments across both platforms.  

Why unified visibility of AI agent risk signals matters

As AI adoption accelerates across Microsoft environments, organizations must manage new forms of behavior, access patterns, and risk. Security teams need more than inventories and permissions. They need visibility into how AI systems operate and how risk evolves over time.  

This integration addresses a critical gap: how to bring behavioral AI security insights into the same control plane as identity, access, and agent management.  

With Darktrace and Microsoft Agent 365 together, organizations benefit from:  

  • Unified visibility: A single pane of glass for understanding AI agent risk signals across Microsoft and Darktrace signals  
  • Faster detection of abnormal agent behavior: Darktrace's Adaptive AI highlights deviations that may not be captured by static controls  
  • Operational efficiency: Security teams can triage and prioritize risk signals without switching between systems
  • Stronger trust in AI deployments: Clear attribution, context, and investigation pathways improve confidence in AI agent usage

Extending Microsoft's AI security model, not replacing it

Securing AI requires a layered approach that combines governance, visibility, threat detection, and risk management. This integration is designed to complement Microsoft's security capabilities, not duplicate them.  

Through Darktrace / SECURE AI, Darktrace contributes:  

  • Identification of risk via advanced prompt analysis  
  • Behavioral anomaly detection across AI agents  
  • Cross-environment threat correlation  
  • Autonomous insight into emerging or unknown risks  

Microsoft provides:  

  • Centralized agent management
  • Identity and access governance
  • Native detection of risk signals and enforcement capabilities

Together, these capabilities create a more complete, layered approach to securing AI-driven enterprises. Organizations gain the governance and policy controls needed to manage AI adoption while benefiting from continuous visibility into how AI is used across the business.

Building the future of secure AI

As AI agents become more deeply embedded in business processes, organizations need more than inventories and static controls. They need to understand how AI is being used, how agents behave, and where risk is emerging across the enterprise.

Darktrace / SECURE AI delivers that understanding through continuous visibility into AI activity, helping security teams assess intent, identify behavioral drift, and uncover emerging risk across both human and agent-driven workflows. Powered by Adaptive AI, it provides the context needed to secure AI as it evolves.  

The integration with Microsoft Agent 365 extends those insights into the workflows organizations already use. Agent 365 provides a unified control plane for governing and securing AI agents, while Darktrace contributes complementary behavioral risk signals that can be surfaced within the Agent 365 experience. Together, they give customers broader context on agent activity and risk while preserving the value of their existing Microsoft and Darktrace investments.

As enterprises move from AI experimentation to AI-powered execution, Microsoft and Darktrace help bring together governance, compliance, behavioral understanding, and oversight in a unified approach to AI security. For organizations adopting Microsoft 365 E7, Darktrace / SECURE AI further strengthens that foundation by providing continuous visibility into AI activity, agent behavior, and emerging risk as AI adoption scales.

Continue reading
About the author
The Darktrace Community
Your data. Our AI.
Elevate your network security with Darktrace AI