AI agents

Injecting dynamic data into prompts safely

• Injecting Dynamic Data into a System Prompt Safely
• Tool Results Are Untrusted: Sanitizing Before Context
• A Precedence Table for Conflicting Sources
• A Tool Result Schema That Carries Trust and Source
• A Trust Field on Every Context Item
• Poisoned Chunks: Sanitizing Retrieved Text Before the Prompt
• Routing Prompts by Data Sensitivity
• Session State Schema: What an Agent Needs Between Turns
• The Context Builder as a Choke Point for Sanitization
• The Folded Summary as an Untrusted Document
• Tool Output Is Untrusted: Treating API Replies as Data