RAG
Prompt injection through retrieved documents: what RovoBlast, GeminiJack and EchoLeak teach RAG teams

Shekhar Kadyan
Co-founder

Prompt injection through retrieved documents happens when an AI assistant fetches a document, email or ticket that contains hidden instructions, and the language model follows them as if the user had typed them. No filter stops this reliably today, so we recommend a retrieval system designed to limit the impact, such as PipesHub, which checks each person's own permissions on every query, connects agents as named users, and cites the exact block behind every answer.
PipesHub publishes this article. PipesHub is an open-source enterprise search product that uses retrieval, so this class of risk applies to it too, and we explain below how its design limits the impact.
Prefer to watch? Here is the five-minute video version of this article.
Key takeaways
Indirect prompt injection means the harmful instruction arrives inside content the assistant retrieves, and the person asking never sees it.
EchoLeak (Microsoft 365 Copilot) and GeminiJack (Google Gemini Enterprise) arrived in 2025 through ordinary content such as an email, a shared document or a calendar invite, while RovoBlast (Atlassian Rovo), fixed on 8 July 2026, arrived through a crafted link.
Per-person permission checks, as in PipesHub, limit what a misled assistant can read.
Block-level citations, as in PipesHub, show exactly which document an answer came from, so an odd answer can be traced to its source.
The defences that hold up are structural: least-privilege access, human approval before actions, limits on outbound links and images, and control over who can add content.
How hidden instructions reach an assistant
Retrieval-augmented generation, or RAG, means that a system fetches relevant passages from company content and gives them to a language model to answer from. Most systems first copy that content into a searchable index through connectors, which are links to apps such as email, Slack or Jira. The model then receives one block of text, called the prompt, which holds the system's rules, the user's question and the retrieved passages together.
A language model has no built-in way to know which part of that text is an order and which part is reference material, so it may obey an instruction that sits in a retrieved passage. OWASP, a non-profit security community, ranks prompt injection first in its 2025 list of risks for applications built on large language models (LLMs). It says the indirect form occurs "when an LLM accepts input from external sources, such as websites or files" (OWASP LLM01:2025).
The person who plants the text needs no access to the assistant, only a way to place content where it will later be read, such as an email, a shared document, a support ticket or a web page.

The path a planted instruction takes through a retrieval system. At step four the model cannot reliably tell data from commands, so per-person permissions, as in PipesHub, limit what it can reach.
The Cloud Security Alliance (CSA), an industry body, wrote in April 2026 that indirect prompt injection "has crossed the line from proof-of-concept to live exploitation" (CSA, 26 Apr 2026).
What the public cases showed
EchoLeak in Microsoft 365 Copilot (June 2025)
Researchers at Aim Labs found that an email sent from outside a company could carry hidden instructions for Microsoft 365 Copilot. When an employee later asked Copilot a routine question, Copilot could retrieve that email along with internal content. Its answer could then include an image whose web address contained internal data, and the image loaded automatically, which sent the data out (arXiv case study, Sep 2025).
Microsoft published the issue on 11 June 2025 as CVE-2025-32711, its public vulnerability identifier. The advisory rates it critical, with a severity score of 9.3 out of 10, credits Aim Labs and marks it as not exploited. It says the vulnerability "has already been fully mitigated by Microsoft" and that users need to take no action (Microsoft advisory). The case study dates the report to January 2025 and the fix to May 2025.
GeminiJack in Google Gemini Enterprise (December 2025)
Noma Labs showed that hidden instructions could sit in a shared Google Doc, a calendar invite or an email. A routine search in Gemini Enterprise could retrieve that content, and the assistant could then place what it found inside an image link that pointed to an outside server (Noma Labs, 8 Dec 2025).
Noma's timeline says it reported the issue on 5 June 2025 and that Google confirmed it on 8 August 2025. Noma says Google separated Vertex AI Search, another Google Cloud product, from Gemini Enterprise and changed how both use their retrieval and indexing systems.
RovoBlast and a second finding in Atlassian Rovo (2026)
RovoBlast arrived by a different route. Varonis Threat Labs found that a crafted link could pre-fill the chat box of Rovo, Atlassian's AI assistant. One click by a signed-in user was enough for Rovo to act on the pre-filled text with that user's access and send what it found to an outside server (Varonis, 7 Aug 2026).
The public bug bounty record shows the report was submitted on 4 November 2025, fixed on Atlassian's servers on 8 July 2026 and disclosed on 14 July 2026. It says the technique was demonstrated against Confluence, Jira and connected apps such as SharePoint (Bugcrowd).
PromptArmor separately reported that hidden text in an uploaded file could lead Rovo to send Jira and Confluence data to an outside web address. PromptArmor says it told Atlassian on 23 May 2026, and its post of 5 August 2026 said Rovo was still vulnerable at that time (PromptArmor).
The post showed no update when we checked on 20 November 2026, so teams that use Rovo should ask Atlassian for the current status. Neither research team reports use against a real customer, a point the CSA also made (CSA, 21 Aug 2026).
Case | How it arrived | How data could leave | Vendor response |
|---|---|---|---|
EchoLeak (Microsoft 365 Copilot) | An outside email | An image address in the answer | Fixed by Microsoft, published 11 Jun 2025 |
GeminiJack (Gemini Enterprise) | A shared document, invite or email | An image link in the answer | Google changed its retrieval setup, says Noma |
RovoBlast (Atlassian Rovo) | A link that pre-filled the chat | A web request by the assistant | Fixed by Atlassian on 8 Jul 2026 |
PromptArmor finding (Rovo) | Hidden text in an uploaded file | A web address the assistant opened | Open on 5 Aug 2026, later status unconfirmed |
The shared pattern: the lethal trifecta
Simon Willison, the developer who coined the term prompt injection, named the pattern "the lethal trifecta" in June 2025. An assistant is exposed when it combines three things. It has access to private data. It reads untrusted content, which he defines as "any mechanism by which text (or images) controlled by a malicious attacker could become available to your LLM". It can also communicate with the outside world (Simon Willison, 16 Jun 2025).
Every case above had all three. Willison also writes that "we still don't know how to 100% reliably prevent this from happening". The practical lesson is to remove or narrow one of the three wherever possible.
What permission checks limit
Permission-aware retrieval means the assistant only fetches what the person asking is allowed to see. It sets the upper limit on what a misled assistant can read. In PipesHub, access is checked when the query runs, against each source's own permissions, for the specific person asking, so a planted instruction can only reach what that person could already open. Varonis wrote that RovoBlast needed "no permission bypass", because Rovo acted with the access of the user who clicked.
Permissions also do not decide who can write into the index. A customer can file a ticket, and anyone can send an email, so content that the user is entitled to read can still carry someone else's instruction.
Defences that help
No single measure is enough. Microsoft describes its own approach as layered, with some defences that reduce the odds of an attack and others that guarantee a particular attack will fail (Microsoft, 29 Jul 2025).
Treat retrieved text as data
The system should mark retrieved passages clearly as reference material, separate from the user's request. Microsoft calls its version of this "Spotlighting". Google adds classifiers, which are models trained to spot injected instructions, and extra security wording around the content (Google, 13 Jun 2025). PipesHub marks text fetched from the web as untrusted, so the model is told to treat it as information, not instructions (our web search guide). These measures lower the success rate, so they should never be the only layer.
Give the assistant the least access it needs
This principle is called least privilege. An assistant should run with one person's permissions, never with a shared account that sees everything. In PipesHub, agents connect as a named user, not as a shared service account. Sensitive areas such as HR and legal can be left out of scope, and unused connectors can be disconnected.
Ask a person before actions
An assistant acts through tools, which are functions it can call to send an email, delete a record or post a message. Such actions should need a person's approval, and OWASP and the CSA both recommend a human check before high-impact actions. In PipesHub, a tool whose action cannot be undone, such as sending an email or deleting data, asks first with an approval card (our toolsets guide). A read-only assistant limits most of the harm to a data leak.
Limit outbound links and images
In EchoLeak and GeminiJack the data left through an image address. The dependable control is to stop answers from loading images or links from domains that are not on an approved list. Google says its sanitiser, a filter that cleans an answer before it is shown, "identifies external image URLs and will not render them". A tool that fetches web addresses needs the same limit.
Control who can add content
Teams should know which sources let outsiders write into the index, such as inbound email, public ticket forms, drives shared with guests and web pages. Each can be indexed separately, labelled, or kept away from assistants that can take actions.
Log and test
The system should record each retrieval, tool call and outbound request, so that odd behaviour can be traced later. PipesHub answers cite the exact block of their source, so a reviewer can trace an odd answer to its document. OWASP recommends regular adversarial testing, which means running controlled attack simulations against your own setup. A harmless test document that asks for a marker phrase is enough to learn whether the assistant follows planted text.
Questions to ask any vendor
Whose permissions does the assistant use when it retrieves?
How is retrieved text separated from instructions in the prompt?
Can answers or tools send data to domains that are not on an approved list?
Which actions need a person's approval, and what is logged for each query?
Where PipesHub fits
PipesHub is built so that a planted instruction reaches as little as possible. It is open source under Apache 2.0, and its design limits the impact of injection in four ways.
Access is checked when the query runs, against each source's own permissions, for the specific person asking. This limits what an injected instruction can reach to what that person can already see.
Agents connect as a named user, not as a shared service account, so an agent sees only what its user can see.
Answers cite the exact block of the source document, so a reader can see exactly which document an answer came from.
Tools whose actions cannot be undone ask first with an approval card, and in the Enterprise edition an admin can set approval rules for each tool.
The code is public and can be self-hosted, so a security team can read how these controls work and run its own injection tests. The Enterprise edition adds audit logs. No retrieval system removes this risk on its own, and PipesHub's design aims to keep the impact narrow.
A team that works entirely inside Microsoft 365 and wants Microsoft's own detection tools, such as Prompt Shields, may prefer Copilot.
To get started, sign up for PipesHub Cloud, or self-host PipesHub with our docs, which also cover toolsets and approvals.
FAQ
What is indirect prompt injection?
Indirect prompt injection is an attack where instructions for an AI assistant are hidden in content that the assistant reads later, such as a document, an email or a web page. The user asks a normal question and never sees the planted text.
Can prompt injection be fully prevented in a RAG system?
No known method prevents it completely. Filters and careful prompt design only reduce the success rate. The dependable approach is to limit what the assistant can read, which actions it can take without approval, and where it can send data.
Do permission-aware search and access controls stop prompt injection?
They do not stop it, but they limit the damage. A misled assistant can only reach what the person asking can already reach. Without per-user permissions, one planted document could expose content from the whole company.
Were EchoLeak, GeminiJack or RovoBlast used in real attacks?
Microsoft's advisory for EchoLeak marks it as not exploited. Noma's GeminiJack post describes research findings and does not report real-world use. Neither Rovo research team reports use against a real customer.
Why are images and links in AI answers a risk?
A web address can carry data inside it. If a misled assistant builds an image address that contains private text, loading that image sends the text to whoever runs the server. Blocking images and links to unapproved domains closes this channel.
Related reading
Sources
Microsoft Security Response Center: CVE-2025-32711, M365 Copilot Information Disclosure Vulnerability, the vendor advisory for EchoLeak, checked 20 Nov 2026.
arXiv: EchoLeak, the first real-world zero-click prompt injection exploit in a production LLM system, an academic case study with the timeline, checked 20 Nov 2026.
Noma Labs: GeminiJack, the researchers' disclosure and timeline, dated 8 Dec 2025, checked 20 Nov 2026.
Varonis Threat Labs: RovoBlast, the researchers' write-up, dated 7 Aug 2026, checked 20 Nov 2026.
Bugcrowd disclosure record for the Rovo link finding, with report, fix and disclosure dates, checked 20 Nov 2026.
PromptArmor: Atlassian Rovo exfiltrates data, the document-based finding and its timeline, dated 5 Aug 2026, checked 20 Nov 2026.
Cloud Security Alliance: Atlassian Rovo prompt injection research note, dated 21 Aug 2026, checked 20 Nov 2026.
Cloud Security Alliance: Indirect prompt injection goes operational, dated 26 Apr 2026, checked 20 Nov 2026.
Simon Willison: The lethal trifecta for AI agents, dated 16 Jun 2025, checked 20 Nov 2026.
OWASP: LLM01:2025 Prompt Injection, definition and mitigations, checked 20 Nov 2026.
Microsoft: How Microsoft defends against indirect prompt injection attacks, dated 29 Jul 2025, checked 20 Nov 2026.
Google: Mitigating prompt injection attacks with a layered defense strategy, dated 13 Jun 2025, checked 20 Nov 2026.
PipesHub docs: web search and toolsets pages, checked 20 Nov 2026.








