AIP-13: Prompt Injection Defense

Abstract#

Defines mandatory injection mitigation model.

Normative Requirements#

1. External content MUST be marked tainted/untrusted.

2. Secrets MUST NOT be injected into LLM context.

3. HIGH-risk actions triggered by tainted content MUST escalate to HIGH.

4. Runtime MUST enforce boundary wrapping for external content.

5. Runtime SHOULD implement two-pass validation for high-risk actions.