LLM guardrails: stopping prompt injection | AI system design

The model will follow instructions hidden in user input unless you wrap it in guardrails.

Wrap the model: filter inputs before it, keep secrets and tools out of its reach, and scan outputs for leaks before anything is sent.

An LLM can't fully tell 'your rules' from 'the user's text.' So a user, or a poisoned document it reads, can inject instructions that hijack it, leak secrets, or make it misbehave.

How you defend it:
- Filter and classify inputs before the model, catching obvious injection and abuse
- Keep secrets and powerful tools out of reach, so even a hijacked model can't do real damage
- Check the output too, scanning for leaked prompts, PII, or policy violations before sending

The catch: guardrails add latency and cost on every call, and they're never perfect. Too strict and you block real questions, too loose and something slips through. It's defense in depth, not a fix.

Heavy input and output filtering, or lock down what the model can actually do?

#systemdesign #techinterview #softwareengineering #codinginterview #faang #interviewprep #ai #llm #security #promptinjection #safety

Видео LLM guardrails: stopping prompt injection | AI system design канала The System Round