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:
- Claude generated a prompt injection payload.
- Hidden instructions were embedded in an email.
- The email passed traditional validation checks.
- Darktrace analyzed the language and sender behavior.
- 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:
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.
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.
