Google and partners dismantle XCPE botnet in more than 1.3m TV boxes

The accumulated papers establish that agent memory works best when it is explicitly curated rather than passively accumulated: updated permission records outperform raw histories, selective forgetting and compressed representations beat full transcripts, and structured episodic stores improve task performance across mobile, robotic, and multi-agent settings. Most recent work has moved toward memory safety and scale, with benchmarks from Tencent and Zhejiang Sci-Tech showing that planted instructions survive into fresh sessions and that proactive retrieval produces 87 to 100 percent false alarms as stores grow, while MIT Media Lab finds push-based messaging collapses under load.

Recent papers from KAIST AI/Microsoft Research Asia, CASIA/UC Berkeley, Alibaba DAMO Academy, and Tencent/NUS collectively establish that video prediction serves as a robust reasoning scaffold: models that explicitly forecast object motion, camera transitions, or scene dynamics consistently improve performance on 3D question answering, navigation planning, physical reasoning, and real-time game control. The line of inquiry has moved from static video understanding to interactive, agentic frameworks where a model either requests predictions on demand (RynnUnify) or maintains explicit game state proxies (Wizard), making video prediction a controllable reasoning step rather than a downstream output.

The accumulated work establishes that world models are most useful to robots as decision-time components rather than pure predictors.