Safety · Privacy · Long-horizon agents

Safe intelligence,
wherever it acts.

As AI becomes embedded across everyday life, we study how intelligent systems can act safely, protect privacy, and remain reliable over long horizons, wherever they operate.

Research Branch

01

Attack & Defense

From adaptive, locally benign attacks, to correctly timed intervention, to agentic safety that reasons over state and harm-enabling actions.

Preprint

SEAD

A State-Based Perspective on Attack and Defense in Tool-Using Agents

Xinjie Shen*, Junran Wang*, Rongzhe Wei, Pan Li

Detect when accumulated agent actions make a later step harmful.

ICML · AI Wild

TurnGate

Learning When to Intervene Against Multi-Turn Malicious Intent

Xinjie Shen*, Rongzhe Wei*, Peizhi Niu*, Haoyu Wang, et al.

Learn when to intervene before multi-turn intent becomes actionable harm.

ICML

CKA-Agent

Bypassing Commercial LLM Guardrails via Harmless Prompt Weaving and Adaptive Tree Search

Rongzhe Wei*, Peizhi Niu*, Xinjie Shen*, Tony Tu, et al.

Adaptively weave harmless-looking prompts into a harmful objective.

02

Physical Privacy

As LLMs become the brains of embodied agents, we study whether they can recognize privacy from physical and social context and turn that understanding into safe actions.

NeurIPS

ImmersedPrivacy

How Far Are VLMs from Privacy Awareness in the Physical World?

Junran Wang*, Xinjie Shen*, Zehao Jin*, Pan Li

Test whether VLMs turn multimodal privacy cues into safe actions.

ICLR

EAPrivacy

Measuring Physical-World Privacy Awareness of Large Language Models

Xinjie Shen, Mufei Li, Pan Li

Measure privacy reasoning across physical, task, and social context.

03

Long-Horizon Agents

We evaluate what agents can sustain over long horizons and scale agentic RL training with verifiable rewards.

Qwen Technical Report

VHD-Play

Generating Agentic RL Environments from Solved Mechanisms

Xinjie Shen, Wei Fan, Xudong Guo, Jianhong Tu, et al.

Scale agentic RL training through environments with verifiable rewards.

Qwen Technical Report

E-CommerceBench

Evaluating LLM Agents on Long-Horizon Autonomous Business Operation

Wei Fan*, Xinjie Shen*, Xudong Guo, Jianhong Tu, et al.

Measure long-horizon capability through sustained autonomous operation.

* Equal contribution.