Prompt Injection Survival Guide for Builders

A short, practical checklist to keep your agent endpoints from becoming a screenshot of shame.

Guardrails

• Sandwich prompts: system preamble + postamble that cannot be overridden.
• Strict tools: validate arguments; whitelist functions and schema.
• No raw tool output to users; sanitize and summarize.
• Rate limit + auth every agent endpoint; log inputs/outputs for forensics.

Quick detections

• High-perplexity inputs; known jailbreak keywords.
• Unusually long tool arguments; base64 blobs; obfuscated code.
• Sudden role shifts (“you are now…”).

Response plan

• Fail safe: refuse or require human review on suspicious inputs.
• Strip instructions from user content before reuse.
• Regression tests for jailbreak prompts before each deploy.

TL;DR

Assume users (or their docs) will try to hijack your agent. Constrain tools, validate input, and log everything.