The Misattribution Gap: When Memory Poisoning Looks Like Model Failure in Agentic AI Systems

This feed contains multiple security-relevant AI research contributions and new attack/defense classes with practical implications for agentic systems, model deployment, and privacy. Key findings: (1) Misattribution Gap / Semantic Norm Drift (SND): memory-layer poisoning of shared vector stores can make agents follow injected policy-like documents across sessions while attribution systems and safety classifiers mislabel the cause as model failure; across 64 documented failures defenders blamed models, four classifiers (including one trained on memory poisoning) produced zero detections across