Hidden in Thought: Transferable Chain-of-Thought Artifacts Induce Harmful Behavior

Collection of 2026 arXiv papers demonstrating multiple practical and systemic AI/ML security and privacy risks plus related secure-systems work. Key findings: (1) “Hidden in Thought” shows harmful chain-of-thought (CoT) traces and distilled reasoning patterns can be transferred or distilled to create highly effective black‑box jailbreaks (raising harmful responses >80% on many open models and improving attacks on strongly aligned models by up to 10×, including GPT‑4.1); (2) FLINT shows 5G PHY-layer scheduling metadata leaks are sufficient to fingerprint federated learning model architecture (F