causal-armor

Defend against indirect prompt injection attacks by detecting dominance shifts using leave-one-out attribution, enabling selective sanitization without sacrificing latency or utility.

Independent GitHub snapshot · Automated classification · Original documentation governs