👋 About

I am a PhD student at the University of California, Santa Cruz, advised by Prof. Chenguang Wang and very fortunate to closely work with Prof. Dawn Song. I interned at Meta Superintelligence Labs in summer 2026 working on post-training agents, and previously worked at Scale Labs on computer-use. I completed my Bachelor’s degree in Data Science from the Mathematics Department at Washington University in St. Louis, graduating with Highest Distinction. You can find my full CV here.

My research focuses on developing methods to better interpret and ensure the safety and performance of large language models (LLMs) and LLM agents.

Agentic AI Safety & Security Interpretability

📣 News

• Sep 2026
  3 works accepted to NeurIPS 2026, and our third Agents in the Wild workshop will be held at NeurIPS in Sydney!
• Jul 2026
  1 work accepted to COLM; SafeClawArena, ChainWorld, and Agent-LE released on arXiv!
• May 2026
  3 works accepted to ICML 2026!
• Apr 2026
  Peer preservation and agent security formalization now out!
• Jan 2026
  RepIt is accepted to ICLR 2026!
• Dec 2025
  Our workshop on Agent Safety is accepted to ICLR - see you all in Brazil! I will be in NeurIPS this week as well.
• Sep 2025
  RepIt and SteeringSafety are now on arXiv!
• Aug 2025
  I will be moving with my advisor to UCSC to continue my PhD! Additionally, AgentVigil was accepted to EMNLP 2025!
• Mar 2025
  Excited to share COSMIC accepted to ACL 2025!

📚 Publications

Preprints

• arXiv

  Agents Interpretability
  Tunable Tool-Call Rates in LLM Agents via Representation Steering

  Y. Chen, V. Siu, Y. Liu, D. Song, C. Wang

  Paper
• arXiv

  Agents Safety & Security
  Understanding and Evaluating Claw-like Agent Security Through a Computer-Systems Lens

  P. Niu, W. Qu, S. Gu, T. Shi, Y. Li, A. Tawaha, H. Alzahrani, V. Siu, B. Li, C. Wang, J. Zhang, B. Alomair, M. Jin, M. Chen, C. Wang, C. Spanos, D. Song

  Paper
• arXiv

  Agents
  ChainWorld: Composing Long-Horizon Desktop Workloads from Atomic OSWorld Tasks

  V. Siu, M. Sharma, D. Song, D. Y. Zhang, C. Wang

  Paper
• arXiv

  Agents Safety & Security
  A Framework for Formalizing LLM Agent Security

  V. Siu, J. He, K. Montgomery, Z. Wang, N. Gong, C. Wang, D. Song

  Paper

Published

• NeurIPS 2026

  Agents Safety & Security
  SudoBench: A Contextual Authorization Benchmark for LLM Agents

  V. Siu, T. Shi, S. Gu, Z. Wang, D. Song, C. Wang
• NeurIPS 2026

  Agents Safety & Security
  Position: A Safe LLM and a Safe Harness Do Not Make a Safe Agent

  V. Siu, K. Montgomery, Y. Potter, Z. Wang, D. Song, C. Wang
• NeurIPS 2026

  Agents
  Agents' Last Exam

  Y. Sun, X. Han, W. Zhang, et al., V. Siu, et al.
• COLM 2026

  Safety & Security Interpretability
  Component and Dimension Sparsity in Transformer Refusal Mechanisms

  V. Siu, G. Grant-Richards, V. Pavlovich, Y. Sun, D. Song, C. Wang
• ICML 2026

  Agents Safety & Security
  Position: Agent Security Needs Redefinition through a Holistic Framework

  V. Siu, J. He, K. Montgomery, Z. Wang, C. Wang, D. Song
• ICML 2026

  Agents Safety & Security
  Peer-Preservation in Frontier Models

  Y. Potter*, N. Crispino*, V. Siu, C. Wang, D. Song

  Paper
• ICML 2026

  Safety & Security Interpretability
  SteeringSafety: A Systematic Safety Evaluation Framework of Representation Steering in LLMs

  V. Siu*, N. Crispino*, D. Park, N. W. Henry, Z. Wang, Y. Liu, D. Song, C. Wang

  PaperCodeData
• ICLR 2026

  Safety & Security Interpretability
  RepIt: Steering Language Models with Concept-Specific Refusal Vectors

  V. Siu, N. W. Henry, N. Crispino, Y. Liu, D. Song, C. Wang

  Paper
• EMNLP 2025 Findings

  Agents Safety & Security
  AgentVigil: End-to-End Redteaming of Black-Box AI Agents

  Z. Wang, V. Siu, Z. Ye, T. Shi, Y. Nie, X. Zhao, C. Wang, W. Guo, D. Song

  Paper
• ACL 2025 Findings

  Safety & Security Interpretability
  COSMIC: Generalized Refusal Direction Identification in LLM Activations

  V. Siu, N. Crispino, Z. Yu, S. Pan, Z. Wang, Y. Liu, D. Song, C. Wang

  PaperCode

* indicates equal contribution.

📰 Media Coverage

🤝 Service

Agents in the Wild: Safety, Security, and Beyond workshop series

NeurIPS 2026 Core Organizer Sydney, Australia ICML 2026 Workshop Staff Seoul, South Korea ICLR 2026 Co-Organizer Rio de Janeiro, Brazil

Featured in Agentic AI Weekly by Berkeley RDI. Sponsored by Scale AI, Skywork, Lambda, AG2, Snorkel AI, Donut Labs, and Aveo Research Labs.

Open Source