6 min read
Defensive Techniques When prompt injection succeeds, unconstrained LLM outputs give attackers unlimited action space. Output schema enforcement and grammar-constrained generation shrink that surface meaningfully, even when injection itself can't be fully prevented.