Agent security
REALM
Identifying Minimal Task Assets for LLM Agents
Identifying the minimal assets an AI agent needs for a user task, addressing overly broad container permissions and resisting malicious instruction injection.
I'm a Ph.D. student at Tsinghua University, working with Associate Professor Long Wang in the Network and System Reliability and Security Lab. My research spans systems security, container isolation, and kernel performance.
Recently, I've been exploring security isolation for AI agents: defining task-specific boundaries for the assets an agent may access.
Agent security
Identifying Minimal Task Assets for LLM Agents
Identifying the minimal assets an AI agent needs for a user task, addressing overly broad container permissions and resisting malicious instruction injection.
Container security · eBPF
No Touching Others' Assets: Detecting Container Escapes via CAGEGuard
Modeling container boundaries through asset ownership and detecting escapes with eBPF. Achieved a 90% detection rate with low false positives and overhead.
Container security · Root cause analysis
Evidence-Guided Root Cause Analysis for Container Escapes
Combining self-correction with agent-oriented log and code search to locate the source of container escapes. Achieved 48% end-to-end function-level accuracy.
Systems performance
A systematic measurement of eBPF performance overhead.
Research evaluation
Picking Valuable Research from a Sea of Citations
Using neural networks to assess citation importance and to judge the values of research works.
Intelligent Technology and Systems Lab, Tsinghua
User click models and reinforcement learning for ranking tasks, advised by Professor Yiqun Liu. Sep. 2022 — Mar. 2023.
Institute of HCI and Media Integration, Tsinghua
Citation importance and research evaluation, advised by Professor Jiangtao Wen. Sep. 2021 — Jun. 2022.
A few things I've built, from research experiments to databases.
Apr. 2023 — Present
Tsinghua University · Advisor: Assoc. Prof. Long Wang
Research on container escape detection, root cause analysis, eBPF overhead, and security isolation for AI agents.
Sep. 2025 — May 2026
Compiler engineering intern
RISC-V compiler development and performance optimization. Implemented the translation pipeline for 1,000+ custom instructions, and developed multiple hardware-aware optimizations.
Jun. — Jul. 2023
Network protocol research intern
Built feature extraction and classification methods for unknown network protocols, using clustering, CNNs, and LSTMs on noisy network traffic.
2024 — 2029 (expected)
Institute for Network Sciences and Cyberspace
GPA: 4.0 / 4.0
2020 — 2024
Department of Computer Science and Technology
GPA: 3.8 / 4.0
Tsinghua University Scholarship for Excellence in Science and Technology Innovation.
In the classroom, and on the slopes.
Undergraduate Compilers Course, Tsinghua
Tsinghua University Ski Team
Tsinghua University Student Ski Association
C / C++ · Python · Rust · LLVM · eBPF · perf · Docker · PyTorch
Keep in touch
On systems, security, or an interesting idea.
renzihou2012@163.com rzh24@mails.tsinghua.edu.cn