EMNLP 2026 · Findings
Hallucination as Exploit:
Evidence-Carrying Multimodal Agents
Requiring verifiable evidence before an agent takes a privileged action.
Guijia Zhang
I study how AI agents can act reliably.
My research focuses on agent security, multimodal reasoning, and the systems that connect model decisions to real actions.
I work on evidence-based authorization, runtime safety auditing, and the evaluation of tool-using agents. I am interested in turning failures observed in working systems into questions we can test.
Selected research
All publications →Hallucination as Exploit:
Preprint · 2026
Do GUI Agents Believe Their Eyes?
Testing how GUI agents use pixels and structured state when the two disagree.
Research software
More projects →FarField
A local research loop that turns a topic into a falsifiable hypothesis and a reproducible experiment protocol. Model proposals are checked with ordinary Python.
OpenClaw Security Suite
Audit logging, policy-based skill isolation, and alerting for an agent gateway, with a dashboard for inspecting security events.
Recent notes
- ECA (Hallucination as Exploit) and STARS accepted to EMNLP 2026 Findings.
- Released FarField, a local research loop for falsifiable hypotheses and reproducible protocols. GitHub
Earlier notes
- New preprint on agent skill regulation as a procedural knowledge representation problem, joint work with HKUST.
- Started a first-author project on evidence-carrying multimodal agents, formalizing hallucination-to-action conversion.
- Completed STARS, an invocation-time risk-auditing framework for LLM agent skills, together with the 3,000-record SIA-Bench benchmark.
- Joined the HKUST Academy of Interdisciplinary Studies as a remote research intern, advised by Prof. Huamin Qu and Yuzhe Shi.
- Began a research collaboration with KAUST (Prof. Di Wang, Shu Yang) on agent runtime safety.
- Joined HKUST (Guangzhou) as an exchange student, working with Prof. Lei Chen on agent memory and database systems.