Security and Human-Centered Assessment of BACnet-Controlled DALI Infrastructure in an Educational Building Automation Testbed

This feed aggregates multiple 2026 security research papers with several high-impact findings. Most critical is “Loss Landscape Poisoning,” which demonstrates a poisoning technique that induces targeted extraction of unseen training records (high extraction rates across language and vision-language models) and shows existing defenses including differential privacy can be bypassed by a follow-on probing attack — a major privacy risk for models trained on sensitive data. Complementary papers highlight operational data-leakage risks in tool-using LLM agents (non-adversarial leakage across common