01
Agentic AI safety & security evaluation
Evaluating the safety of tool-using language-model agents. Ongoing work is presented at the topic and status level only.
Research / Public scope
Safety evaluation
across models,
tools, and workflows.
My work asks how technical risk can be turned into structured evidence, how that evidence should be evaluated, and how findings can become controls that people can review and use.
- Research themes
- 03
- Public case studies
- 04
- Disclosure
- Public scope only
Current focus / July 2026—present
Safety evaluation of tool-using LLM agents.
Ongoing independent research on safety evaluation for tool-using language-model agents.
Disclosure boundary
Only the title, role, period, location, and general topic are public. Methods, benchmarks, models, prompts, traces, code, data, findings, and manuscripts are not disclosed.
Research lines / Connected, not scattered
Three ways into the same problem.
Each line asks how evidence can make AI-system safety more legible, repeatable, and operational.
02
Safety testing & assurance
Turning public risk principles and technical failure models into structured tests, evidence, deployment decisions, and mitigation checks.
03
Multilingual content & model safety
Building evaluation workflows for toxic, xenophobic, misleading, and biased content across multilingual settings and document formats.
Working approach / Public across completed projects
From risk question to operational evidence.
This is a public synthesis of methods used across completed work. It is not a description of the private design of the ongoing agent-safety study.
01
Frame the risk
Translate a broad safety concern into explicit system boundaries, failure categories, evaluation questions, and decision criteria.
Threat models · Risk taxonomies · Standards mapping02
Design the evaluation
Connect test cases, model or system behavior, and multidimensional measures so findings can be reproduced and compared within a defined scope.
Adversarial testing · Experimental design · Evaluation metrics03
Structure the evidence
Build data, classification, and review workflows that make model outputs legible to researchers, operators, and decision-makers.
NLP pipelines · Human review · Failure classification04
Operationalize the result
Carry evidence into safety cases, mitigation verification, product requirements, and workflows that can be used beyond a single analysis.
Safety cases · Control checks · Workflow designApplied systems / Public case studies
Where research structure becomes a tool.
01AI Safety Testing Framework for LLM and Agentic SystemsAI safety assurance↗02AI-Driven Media Analysis for Toxic and Xenophobic Language DetectionMultilingual content safety↗03Forecasting the Term Structure of Interest Rates with Dynamic SV ModelsQuantitative research↗04B2B Digital Operations PlatformProduct & systems delivery↗
Springer conference paper / 01
Springer conference paper
The Determinants of Carbon Emissions in Macau: Based on the Analysis of STIRPAT, EKC, and LMDI Models
Wenlan Liang, Ruizhen Pan, Yilin Cao, Wentong Wang, Shixuan Lin, and Yujia Zhao
IEIS 2022, Lecture Notes in Operations Research, pp. 152–161. Springer Nature Singapore.
DOI 10.1007/978-981-99-3618-2_15External record only
This site links to the DOI record and does not reproduce or host the paper.
Open channel / Collaboration