Jun Sun

This survey provides a comprehensive, structured review of safety research in Embodied AI, analyzing attacks and defenses across the entire embodied pipeline to guide the development of safe, robust, and reliable real-world agents.

The paper proposes SafeClaw-R, a novel framework that enforces safety as a system-level invariant over the execution graph to mitigate the high safety and security risks inherent in autonomous multi-agent LLM systems.

ClawGuard is a novel runtime security framework that deterministically enforces user-confirmed rules at tool-call boundaries to protect LLM agents from indirect prompt injection.

The paper introduces Salami Slicing Risk, a novel multi-turn jailbreak technique that accumulates harmful intent through numerous low-risk inputs, achieving state-of-the-art attack success rates against major LLMs.

FunPoison introduces a functionality-preserving poisoning technique that injects small, compilable weak-use fragments into code datasets to prevent unauthorized use of CodeLLMs without breaking the code's functionality.

The paper introduces Layerwise Convergence Fingerprinting (LCF), a tuning-free runtime monitor that detects various LLM misbehaviors (backdoors, jailbreaks, prompt injections) by analyzing the trajectory of hidden states between layers.

The paper introduces Dr-CiK, a new benchmark designed to evaluate agents' ability to proactively discover, filter, and utilize relevant external context for time series forecasting, demonstrating that current agents struggle significantly with this task.

MobEvolve introduces an agentic self-evolving heuristic system that significantly improves human mobility generation by iteratively refining its internal logic using an LLM agent, outperforming deep generative and LLM-based methods.