How the attack works
An attacker creates or modifies a document that looks legitimate: a business report, a job resume, a legal snippet, a factual note, or Whois registration data. Inside that document they insert a concealed instruction, using phrases like 'ignore the above and...', 'your task is now to...', an HTML comment such as '<!-- DO THIS INSTEAD: ... -->', or an encoded command hidden in base64 or code-comment syntax. The document is then submitted to an AI agent for a normal task, such as summarizing, reviewing, or extracting information. Because the agent reads the whole document as input, it can pick up the hidden instruction and treat it as a legitimate command, deviating from what the user actually asked for.
Why it matters
An organization can lose control over what its AI agent does next — the agent may leak data, call unintended tools, or carry out actions the attacker chose, all triggered by a document that looked harmless.
What you can do
- Treat all document content passed to an agent as untrusted input, not as instructions, and enforce this separation in how prompts are constructed.
- Scan uploaded or ingested documents for suspicious patterns like hidden HTML comments, base64 blobs, or phrases such as 'ignore the above' before passing them to an agent.
- Restrict what actions or tool calls an agent can trigger based solely on content extracted from a document, especially reports, resumes, or third-party data feeds.
- Log and review cases where an agent's output diverges sharply from the user's original request after processing a document, since this can indicate a successful hijack.
Known benign look-alikes
- Security audit reports that quote injection examples inside the document body
- Red team exercise documents containing injection test cases as examples
- Legitimate documents that coincidentally contain instruction-like language