Skip to content
Get in touch

Reviewed in POC · content version 2

Review Prompt Injection

Your draft is bound to this exact source hash. If the term changes, the Review Desk will use a new storage key instead of silently attaching old decisions.

Definition

Prompt injection is an attack in which instructions placed in user input or retrieved content attempt to override, redirect, or exploit an AI system's intended instructions and tool access.
Decision for Definition

Plain-Language Explanation

A model may encounter text that looks like a new command inside a document, webpage, email, or direct request. If the surrounding system treats that content as trusted instruction, the model may follow it.
Decision for Plain-Language Explanation

Why It Matters

Systems that read untrusted content or use tools can expose data or take unwanted actions unless permissions and trust boundaries are enforced outside the prompt.
Decision for Why It Matters

Example

A document contains hidden text telling an assistant to ignore its task and send retrieved confidential information to an external address.
Decision for Example

Common Misconception

Prompt injection cannot be solved reliably by adding one stronger sentence to the system prompt. It requires architectural controls, limited permissions, isolation, and monitoring.
Decision for Common Misconception

Remember It This Way

Untrusted content tries to become trusted instruction.
Decision for Remember It This Way

Related Terms

* [[system-prompt|System Prompt]] * [[adversarial-attack|Adversarial Attack]] * [[guardrails|Guardrails]] * [[sensitive-information|Sensitive Information]] * [[assumed-isolation|Assumed Isolation]]
Decision for Related Terms

Sources

* [NIST AI 100-2 E2025, Adversarial Machine Learning](https://doi.org/10.6028/NIST.AI.100-2e2025) — authoritative taxonomy and definition of prompt-injection attacks.
Decision for Sources

Learn More

* Treat retrieved content as untrusted input and preserve boundaries around [[sensitive-information|Sensitive Information]].
Decision for Learn More
Return to term

Not saved yet.