Prompt Injection Defense

Harden LLM agents and autonomous tool execution chains against direct jailbreaks and indirect prompt injection attacks by enforcing bounded tool parameter schemas, structural content delimiters, and canary token validation.

When to Use

• Ingesting untrusted external data (web pages, third-party emails, customer documents, database records) into agent prompts
• Exposing state-changing tools (APIs, filesystem writes, email dispatches, database mutations) to LLM agent workflows
• Designing multi-agent communication protocols where downstream agents process upstream untrusted messages
• Auditing agent system prompts and execution boundaries against OWASP ASI01 (Goal Hijack) and ASI02 (Tool Misuse)
• Verifying safety controls before promoting agent capabilities into production environments

Core Rules