How to Defend LLM-Powered Apps Against Prompt Injection Attacks

Prompt injection is a security vulnerability where users manipulate AI-powered applications by typing instructions that override the developer's original system prompt. Because large language models treat all text equally, they cannot inherently distinguish between trusted developer instructions and untrusted user input. Developer Maneshwar, who builds an open-source AI code review tool called git-lrc, outlines several practical mitigation techniques including input filtering, inline security warnings, and post-prompting. Strategies such as 'sandwich prompting'—placing user input between two sets of instructions—can make injection attacks harder to execute without requiring complex infrastructure. While no single method eliminates the risk entirely, combining these lightweight defenses significantly raises the bar for would-be attackers.