新的 AI 安全工具 Arc Sentry 狂勝 LLM Guard，檢測率高 92%，對手只有 70%——它到底怎麼做到的？

I built Arc Sentry, a pre-generation prompt injection detector for open-weight LLMs. Instead of scanning text for patterns after the fact, it reads the model’s internal residual stream before generate() is called and blocks requests that destabilize the model’s information geometry. Head to head benchmark on a 130-prompt SaaS deployment dataset: Arc Sentry: 92% detection, 0% false positives LLM Guard: 70% detection, 3.3% false positives The difference is architectural. LLM Guard classifies i