CAITLYN adds a second system that tries to write new injection defenses when it sees an unfamiliar signal

arXiv·medium signalXBlueskyLinkedInCopy link

Injection defenses face a trilemma between runtime efficiency, contextual precision and adaptability, and mostly fail on variants they were not built for. CAITLYN is agent-agnostic middleware with two halves: System I handles known attacks through a Tier-0 library of rule-based detection scripts backed by Tier-1 optimized LLM inference, while System II watches for abnormal signals and attempts to synthesize a new defense for them. It matches standard-benchmark detection performance while adding the synthesis path, which is the part worth watching for anyone maintaining a hand-written injection rule list.