Pengfei He
I am a Research Scientist at Google Cloud AI Research, working on AI security and trustworthy LLM agents. I received a dual PhD in Computer Science and Statistics from Michigan State University, where I was advised by Prof. Jiliang Tang (Computer Science) and Prof. Yuehua Cui (Statistics).
As large language models evolve into autonomous agents that reason, plan, and act through external tools, they bring new kinds of security and reliability risks. My research aims to understand these risks and to build agentic AI systems that can operate safely in the real world:
- Uncovering vulnerabilities. I study how LLMs and agents fail, including jailbreaks (EMNLP 2024), malicious message flow between agents (ACL 2025), agent memory privacy (ACL 2025), and memory injection attacks (NeurIPS 2025). I also build automated red-teaming and evaluation frameworks (EACL 2026) to expose these weaknesses systematically.
- Securing agents. I develop lightweight defenses that use an LLM’s own attention signals to monitor inter-agent message flow, assessing whether each message can be trusted before agents act on it (NeurIPS 2026).
- Reasoning, interpretability, and tool use. I study how models reason (EMNLP 2025, Stat, AISTATS), how their behaviors can be interpreted and controlled (NeurIPS 2026), and how they use tools (ICLR 2026).
My path into this area began with the robustness of deep learning models, through work on adversarial attacks (ICML 2023) and data poisoning (ICLR 2024, TMLR 2024). That work laid the foundation for my current research on LLM and agent security. Looking ahead, my goal is to develop principled foundations and evaluation methods for safe and reliable agentic AI.
I enjoy working across disciplines, especially on deploying AI safely in high-impact domains such as education (TALE 2025, EMNLP 2026), cybersecurity (ICML 2026), and scientific discovery. Feel free to reach out if you would like to chat or collaborate!
When I am not doing research, I am usually being supervised by my two cats, Fisher and Huli (meet them here).
For AGI and Security!
news
| Sep 25, 2026 | 2 papers accepted to NeurIPS 2026! Including To trust or not to trust: Attention-based Trust Management for LLM Multi-Agent Systems and Crafting Reversible SFT Behaviors in Large Language Models. |
|---|---|
| Aug 29, 2026 | I am honored to be invited as the Poster Chair for KDD 2027. |
| May 2, 2026 | My work at Google, Co-RedTeam, has been accepted to ICML 2026! |
| Apr 7, 2026 | 3 papers (2 main, 1 Findings) accepted to ACL 2026! Including agentic memory management, understanding dynamic retrieval heads and dual-phase reward modeling for test-time reasoning. |
| Feb 23, 2026 | I am happy to win the Outstanding Graduate Student Award! |
| Jan 26, 2026 | My work, Traject-Bench, has been accepted to ICLR 2026! |
| Oct 6, 2025 | I am starting a new position as Student Researcher at Google, Cloud AI Research. |
| Sep 27, 2025 | I am proud to share Traject-Bench, the first trajectory-aware agentic tool usage benchmark! |
| Sep 21, 2025 | Our work Memory Injection Attacks on LLM Agents via Query-Only Interaction is accepted to NeurIPS 2025. |
| Aug 21, 2025 | Three works accepted to EMNLP 2025! LLM reasoning, LLM uncertainty, LLM privacy. |
selected publications
-
NeurIPS 2026To trust or not to trust: Attention-based Trust Management for LLM Multi-Agent SystemsIn NeurIPS, 2026
-
ICLR 2026TRAJECT-Bench: A Trajectory-Aware Benchmark for Evaluating Agentic Tool UseIn ICLR, 2026
-
ICML 2026Co-RedTeam: Orchestrated Security Discovery and Exploitation with LLM AgentsIn International Conference on Machine Learning, 2026
-
ACL 2025Red-Teaming LLM Multi-Agent Systems via Communication AttacksIn Association for Computational Linguistics, 2025
-
ACL 2025Unveiling Privacy Risks in LLM Agent MemoryIn Association for Computational Linguistics, 2025
-
ICLR 2024Sharpness-Aware Data Poisoning Attack Spotlight (top 5%)In International Conference on Learning Representations (ICLR), 2024
-
JMLRLarge sample spectral analysis of graph-based multi-manifold clusteringJournal of Machine Learning Research (JMLR), 2023