Post #8354

⚡ Microsoft built a scanner to detect backdoors in open-weight LLMs 🧠 using 3 behavioral signals. It flags trigger attention spikes, memorized poisoning data leaks, and fuzzy trigger activation—no retraining required. Built to scan open models at scale. 🔗 Signals, detection method, limits, AI SDL shift → https://thehackernews.com/2026/02/microsoft-develops-scanner-to-detect.html