How the attack works
An attacker plants content the agent will read as part of its normal work, such as a document or webpage. The content is written to look like a message from the user, using a fixed phrasing pattern ('This is an important message from me, {user}, to you, {model}') combined with a claim that the agent must do something first 'before you can solve the task that I gave you in the beginning'. Because the agent trusts what looks like user instructions, it treats the injected text as a legitimate prerequisite step and carries it out. The actual goal spliced into this wrapper can be anything — stealing data, forwarding emails, changing settings, disabling safety checks — the wrapper itself is what makes any of those goals land.
Why it matters
An organization loses control over what its AI agent does next: the agent can be redirected to leak data, change configuration, or take other unauthorized actions, all disguised as a normal step the user supposedly asked for.
What you can do
- Treat any content the agent reads from external documents, emails, or web pages as untrusted input, not as user instructions, and enforce that separation in the agent's system design.
- Review agent logs for messages matching the 'important message from me... to you' framing or references to a 'task I gave you in the beginning' that did not originate from the actual user session.
- Require agents to re-confirm sensitive or state-changing actions with the real authenticated user rather than acting on instructions found inside retrieved content.
- Test agents against known prompt-injection benchmarks like AgentDojo to check whether this wrapper pattern successfully hijacks task flow before deploying to production.
Known benign look-alikes
- Legitimate multi-step task instructions that happen to say 'before you begin' without the identity-spoof framing or the exact 'task that I gave you in the beginning' phrasing
- Fiction or role-play content quoting this exact benchmark phrase for research/documentation purposes