Skip to content
anythingai.space
AI Securityintermediate12 min · 9 steps

Input Guardrails

Fast, parallel checks on every incoming message that decide allow, redact, redirect or block before the main model runs.

unfilteredWHAT ARRIVES IN ONE AFTERNOONVPN won't connectfinemy card is 4111 …PIIignore your rulesinjectionwrite my essayoff-topicyou useless botabuseMessageevery turnMain modelhandles everything
requesterrorsignal
1 / 9
ProblemStep 1 of 9

The main model sees everything users send

Without a gate, every message goes straight to the expensive main model: personal data, injection attempts, off-topic requests, abuse. The model then has to handle all of it correctly, every time.

Input guardrails are fast checks on every incoming message before the main model: PII, injection, scope and abuse. They decide allow, redact, redirect or block.

What changed
  • +Messages hit the model unchecked

Lesson 7 of 22 in AI Security