Detecting Semantic Camouflage via Latent Intent Verification in SLMs

Research on Phi-3, Qwen2.5, and Gemma-2b identifies an Intent Horizon where harmful intent is masked by benign narratives. The study shows that standard guardrails fail because harmful representations are often hidden deep within the model's latent activation trajectories.