The idea
Input guardrails run before the main model call. They can block or flag injection attempts, off-topic requests, abusive content, personal data that should not be sent, oversized inputs and known-bad patterns.
Use layers: cheap deterministic checks (length, file type, regex for secrets or card numbers), then classifiers (a small LLM prompt or a dedicated model such as Llama Guard or a provider moderation API). Run them in parallel with the main call where possible to save latency, and cancel the main call if a check fails.
Guardrails reduce risk; they do not eliminate it. Measure false positives as carefully as misses.
They stop the cheapest attacks and most accidental misuse before you spend tokens on them.
Words to know
- GuardrailA check that can block, modify or flag a request or response.
Try it without code
List your inputs
- List every way text enters your model (chat, uploads, fetched pages, tool results).
- Write which guardrail applies to each.
Use it in code
Layered input checks, deterministic first.
import re
MAX_CHARS = 8000
SECRET_PATTERNS = [r"sk-[A-Za-z0-9]{20,}", r"AKIA[0-9A-Z]{16}", r"\b(?:\d[ -]?){13,16}\b"]
def check_input(text, classify_injection):
if len(text) > MAX_CHARS:
return False, "Message too long. Please shorten it."
for p in SECRET_PATTERNS:
if re.search(p, text):
return False, "It looks like you pasted a key or card number. Please remove it."
if classify_injection(text): # small LLM or dedicated model
return False, "I can only help with questions about your account."
return True, NoneAdd input guardrails to your agent and run your 20-attack set plus 30 normal messages; report block rate and false positives.
Common mistakes
- Only checking the chat box, not uploaded files and tool results.
- Error messages that teach attackers what was detected.