Agent security

Questions answered

• Can a system prompt stop prompt injection?

  Why instruction wording cannot be the only prompt-injection control, and how to build boundaries around evidence, tools, authority, and effects.

• How should an AI agent tool-permission system work?

  A concrete authorization design for agent tools using identity, scoped capabilities, argument validation, risk tiers, confirmation, and effect receipts.

Primary field guide

Securing AI Agents — Prompt Injection, Tool Permissions, Identity, Sandboxing, Memory Poisoning, and Audit Trails

Related books

• Context Engineering in Production
• Observability for AI Agents

Reusable resources

• Prompt-injection control matrix (Markdown)
• Agent run envelope schema (JSON Schema)