News
2026
November
Invited to give talks at AI Scientists: From LLM Agents to Autonomous Research at DAI 2026 and DeepSearch: 1st Workshop on Deep Search and Research Agents for the Web at WSDM 2027, both in Hong Kong.
October
Recognized as a CVPR 2026 Outstanding Reviewer.
Invited to give talks, “SciencePal 2.0 Self-Evolving Research Agent”, at the Auto-Research Forum of CNCC 2026 in Chengdu and the Medical Large Model Agent Forum of CHIP 2026 in Dalian.
September
Our recent work has been accepted to NeurIPS’26 (Internal Safety Collapse and VEX-Bench) and EMNLP’26 Findings (HarmProfile). Congratulations to all collaborators!
Invited to give a talk, “Trustworthy Recursive Self-Improvement”, at the Southern University of Science and Technology.
August
Invited to give a talk, “Towards Trustworthy Agents: From Red Teaming to Trustworthy Recursive Self-Improving”, at Huawei Hong Kong Research Institute.
June
Our DropVLA, an action-level backdoor attack on vision-language-action models, is accepted to IROS’26. Congratulations to Zonghuan and all collaborators!
Invited to give a talk, “Hands-on with Claude Code: Frontier Code Agents for Research and Browser/Computer Use”, at the 2026 CityUHK Joint Summer School on AI for Science.
Invited to give a talk, “SciencePal 2.0: Your Self-Evolving Science Agent Team”, at the 2026 Tencent Cloud AI Industry Application Conference in Beijing.
Invited to review for ACM Computing Surveys (CSUR) and IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI).
May
Received the ICML 2026 Gold Reviewer Award (top 25% of reviewers).
Our Defense-to-Attack study on jailbreaking vision-language models is accepted by Pattern Recognition (PR). Congratulations to Yunhan and all collaborators!
March
We released System-Prompt-Open, an open database of system prompts extracted from frontier LLMs (project page).
Our survey on embodied AI safety, covering 400+ papers on risks, attacks, and defenses across perception, cognition, planning, interaction, and agentic systems, is now on arXiv.
Joined HKAI-Sci as a Research Assistant Professor.
Our OpenRedRL benchmark for RL-based red teaming is published in Frontiers of Computer Science (FCS).
February
Our GenBreak, red teaming text-to-image generation with LLMs, is accepted by CVPR’26. Congratulations to Zilong and all collaborators!
January
Released Just Ask, a framework where curious code agents reveal system prompts in frontier LLMs.
Our survey Safety at Scale on large model and agent safety is published in Foundations and Trends® in Privacy and Security (FnT P&S). Congratulations to all collaborators!
2025
January
Our BlueSuffix, reinforced blue teaming for vision-language models, is accepted by ICLR’25. Congratulations to Yunhan and all collaborators!
2024
December
Our work on curiosity-driven auditing for LLMs is accepted by AAAI’25.
April
Our work on constrained intrinsic motivation for RL is accepted by IJCAI’24.
March
Our work on adversarial policy learning in RL is accepted by DSN’24.