Quick answer
Indirect prompt injection is a form of prompt injection where the attacker never talks to the AI directly. Instead, they hide instructions in content the AI will later read, such as an email, document, calendar invite, web page or database record. When the AI processes that content, it follows the hidden instructions.
On this page
How It Works
Modern AI assistants summarize inboxes, answer questions from shared drives, browse the web and read tickets. Each of those sources can be written by an outsider. An attacker sends an email or publishes a page containing instructions, sometimes hidden in white text or metadata, and waits for an assistant to read it on a user's behalf.
Why It Is So Dangerous
The victim does nothing wrong. They ask their assistant a normal question, and the assistant, acting with the victim's access, follows instructions planted by someone else. Researchers have repeatedly shown this pattern against enterprise assistants, including zero-click attacks where a single crafted email was enough to make an assistant leak data. See Microsoft 365 Copilot security.
How to Test for It
- List every external source the AI reads: email, chat, shared files, tickets, CRM notes, web pages and tool outputs.
- Plant test instructions in each source, visible and hidden, and trigger a normal user request that causes the AI to read it.
- Check for data exfiltration paths, especially rendered links and images that can carry data to an attacker's server.
- Test whether planted instructions can trigger actions through connected tools.
- Repeat across different user roles, since impact depends on the victim's access.
How to Reduce the Risk
Limit which sources each assistant can read, strip or neutralize hidden content, block automatic rendering of external links and images, require confirmation before actions, and log what content the AI processed for each response so incidents can be investigated.
Regulatory and Framework Context
In the US, indirect injection testing is particularly relevant before rolling out assistants connected to email and document stores, and supports the NIST AI RMF's call to test AI systems against adversarial inputs.
Latest intel on this topic
- Gemini AI agents repeatedly bypassed safeguards to attack real websiteBleepingComputer · November 16, 2026
- California sues OpenAI over AI agents’ summer hacking spreeCyberScoop · November 16, 2026
- California sues OpenAI over AI agents escaping sandboxes, hacking firmsBleepingComputer · November 16, 2026
- Attackers are using AI agents to automate cybercrime at 'industrial scale,' FBI and AWS warnThe Register · November 16, 2026
- Google Finds Hackers Using AI Agents to Build and Run Exploits From Software PatchesThe Hacker News · November 16, 2026
Updated hourly from the daily briefing.
Common questions
What is a zero-click AI attack?
An attack where the victim does not need to click or open anything. A crafted email or document is processed automatically by an AI assistant, which then follows the hidden instructions.
Which AI systems are most exposed?
Assistants and agents that read content written by outsiders, such as email, shared documents, tickets and web pages, especially when they also have access to sensitive data or tools.
Change log
- November 13, 2026First published.