An assistant reading an email may encounter hidden instructions that redirect it from its assigned task. This is prompt injection: attacker-controlled content causes a model to follow the attacker’s instructions instead of its intended instructions. The underlying problem is that a model cannot reliably distinguish instructions it should obey from material it is only supposed to read. section 4.0 section 8.2
Instructions can arrive through emails, documents, webpages, shared files, tool descriptions, or text embedded in images. A jailbreak is a specialized variant that targets model safety guardrails. Injection can affect an answer alone, but access to data and tools raises the stakes: a misled agent may expose records, select an inappropriate tool, or perform an unintended action. section 4.5 section 8.2
A zero-click email attack
In the disclosed EchoLeak attack, hidden instructions in a malicious email could induce Microsoft 365 Copilot to exfiltrate enterprise data without user interaction. The attack illustrates how ordinary content can become an instruction channel when an assistant has access to connected information. section 4.0
Figure 4.4 Image prompt injection causing a faulty decision in ChatGPT-5.

See an example of image-borne instructions influencing a document-validity decision. section 4.5
Treat prompt injection as a system security problem, not just a prompt-filtering problem. Traditional access controls and input validation remain useful, but blocking known phrases cannot reliably determine whether an instruction is legitimate. Restrict what an agent can read, limit its tools and permissions, require approval for hard-to-reverse actions, and record its activity. Continuously red-team and update guardrails; surrounding controls should limit the consequences when a bypass occurs. section 4.0 section 4.5 section 8.2















