Overview

• LLM-based recommender systems leak up to 65% of user interaction histories and infer age and gender correctly in 87% of cases through inversion attacks
• Autonomous LLM agents remain vulnerable to direct prompt injection (94.4%), retrieval-augmented backdoors (83.3%) and inter-agent trust exploits (100%), enabling full system takeover
• Reviews of fact-checking and hallucination evaluation frameworks reveal persistent failures in detecting misinformation across specialized domains
• TRAIL unifies joint inference with dynamic knowledge graph refinement to outperform existing knowledge-augmented baselines by 3%–13%
• Self-reward reinforcement learning and CodeBoost enable LLMs to drive self-improvement using internal reward signals and abundant code snippets without extensive human supervision