LiteLLM Prompt Injection Protection Design
Akto, Thursday, October 22nd, 2026
Akto explains why LiteLLM's keyword-based prompt injection check is easy to bypass and how to configure stronger guardrails.
Akto says LiteLLM's built-in prompt injection check fuzzy-matches a predefined list of verbs and words against requests.
Simple tricks such as Base64-encoding "ignore previous instructions" slip past it because the gateway looks at strings while the attack lives in meaning.
The post explains why pattern filters fail against direct and indirect injection and why intent should be classified by a model.
It walks through the guardrails LiteLLM supports, how to set up blocking in config.yaml and how to test the setup.