A prompt injection can turn externally supplied content into instructions the model follows instead of its intended behavior. Inputs may come from public users, employee-pasted material, or documents ingested by retrieval-augmented generation (RAG) systems. section 4.2 section 4.5
Prompt injection targets the model’s behavior; a jailbreak specifically targets its safety guardrails. The consequences depend on what the application can do: manipulated output may be merely embarrassing, while access to sensitive data or real-world actions can make it a breach or operational failure. section 4.5 section 4.2
Assess the surrounding capabilities, not just the input:
- Untrusted input, sensitive data access, and external communication together form the “Lethal Trifecta”; breaking any one element prevents the described data-leak attack. section 4.6
- For permitted browser agents, use isolated profiles without saved credentials or corporate sessions; logged-out mode reduces capability while removing access to sensitive data. section 4.8.2
- Use enterprise secure-browser solutions to log navigation and agent actions in real time, since traditional endpoint monitoring cannot see inside AI browsers. section 4.8.2
Figure 4.4 Image prompt injection causing a faulty decision in ChatGPT-5.

Illustrates image-embedded prompt injection causing a faulty decision. section 4.5

