The Orchestrator › Vocabulary › Prompt injection
Prompt injection
Instructions that reach the model through content it reads — email, tickets, pages, records — rather than through the system prompt. Direct when the attacker types them; indirect when they arrive as data.
Discussed in the book
- Chapter 5: Keeping Agents Uncompromised — 5.6 The orchestrator’s security job is not Infosec’s job mentioned 3×
- Chapter 2: The Instructions Are Part of the System — 2.4 Learning, traces, and authority mentioned 2×
- Chapter 5: Keeping Agents Uncompromised — 5.1 What compromise looks like mentioned 2×
- Chapter 2: The Instructions Are Part of the System — 2.1 Layers beneath the prompt mentioned once
- Chapter 2: The Instructions Are Part of the System — 2.2 The instruction tree mentioned once
- Chapter 2: The Instructions Are Part of the System — 2.5 Diagnosing the environment mentioned once
- Chapter 4: Living with Nondeterminism — 4.2 The failure envelope mentioned once
Related essays
- No published essay touches this term yet.