AI security
Studying how LLM applications separate trusted tasks from untrusted content, and how defenses affect both security and utility.
Prompt injection / Trust boundaries / LLM applications
Researcher / Builder / Engineer
Junjie Li.李骏杰
Researching and building secure, reliable AI systems.
AI for Security·Harness Engineering·Software & Systems Security
01Research focus
Questions worth working on.
Where security meets the systems we build—and the evidence we use to trust them.
Harness engineering
Building structured datasets, controlled comparisons, and inspectable evaluation workflows for software and AI systems.
Reproducibility / Benchmark design / Empirical evaluation
Software & systems security
Modeling authorization logic and system behavior, then checking whether comparable execution paths enforce consistent rules.
Authorization / Logic vulnerabilities / System behavior
02Selected work
Ideas, put into practice.
A selection of research and engineering work. Built to understand, evaluate, and improve.
SpecGuard
Channel-aware defense for indirect prompt injection in LLM applications.
Research notes
A black-box defense that separates trusted tasks from instructions embedded in untrusted content. The recorded local benchmark covers 1,546 curated records, four models, and six defenses; aggregate ASR changes from 0.671 to 0.322, with a documented utility trade-off. These are local results, not independent replication.
HVCR
A benchmark for causal reasoning in physically grounded video scenarios.
Research notes
Contributed to controlled physics simulation, rendering, and video processing at Shanghai AI Laboratory. The work includes video-question construction and comparative evaluation of multimodal models and human responses; no current submission outcome is asserted.
DCASE / FISHER
Pretrained audio model transfer and fusion for industrial anomalous sound detection.
Research notes
Explored self-supervised audio model transfer, generative audio augmentation, Bayesian-optimized ensembles, and multi-sampling-rate fusion. The available project status is recorded as of December 2025; rankings, publication status, and numerical results are not asserted.
CLIP implementation
Image–text contrastive learning with controlled encoder and training experiments.
Research notes
Implemented a CLIP-style dual encoder in PyTorch with an InfoNCE objective. Built a Flickr8k training pipeline using AdamW, cosine scheduling, and image augmentation, and explored BERT and Vision Transformer encoders. Retrieval metrics remain omitted pending evidence review.
Further explorations
04 PROJECTSReliable industrial sound detection
2026Foundation-model transfer and reliable model selection for acoustic monitoring.
Research notes
Bachelor’s thesis research combining foundation-model transfer, reliable model selection, fusion calibration, and dual-channel evidence for industrial anomalous sound detection.
Logic vulnerability dataset
2026A CVE/CWE-centered corpus for recurring logic-vulnerability scenarios and mechanisms.
Research notes
The reviewed repository snapshot from August 2026 records 2,626 analyzed rows across a 23-CWE allowlist, with a Python pipeline for structured scenario and mechanism reports. The counts describe that snapshot, not a live database.
JS Web authorization research
2026An evidence-first workflow for authorization logic in JavaScript and TypeScript backends.
Research notes
A four-stage research method separating repository selection, safe-anchor discovery, authorization-context recovery, and V1/V2 comparison. It describes a conservative research workflow, not a validated one-command vulnerability scanner.
Robot vision utilities
Camera calibration and screen-based flower recognition for a robotics project.
Research notes
Computer-vision modules using OpenCV, NumPy, and Jupyter. The reviewed repository supports the implementation summary; it does not establish project dates, competition results, or an individual leadership role.
03Research journey
Across places. Through ideas.
A few places that have shaped how I think—and what I work on next.
An evolving map04 PLACES
Selected · Singapore
Drag to explore · Select a place
Singapore
1.35°N / 103.82°E
Beginning the PhD journey
Nanyang Technological University
PhD student in the College of Computing and Data Science, supervised by Prof. Xiaofeng Wang. Research interests include AI for security and harness engineering.
Different coordinates. Connected questions.04 PLACES 05 CHAPTERS
04Experience & education
Learning by doing.
Research experience
2026
Independent research
Authorization logic & LLM application security
Developing evidence-first workflows for logic-vulnerability analysis, historical case organization, and prompt-injection defense evaluation.
Related work2025.06 — 2025.09
Shanghai AI Laboratory
Research intern
Contributed to a causal video-reasoning benchmark and a physics-based data-generation pipeline as a listed collaborator.
Related workEducation
2026.08 — Present
Nanyang Technological University
PhD Student · Computing and Data Science
Supervised by Prof. Xiaofeng Wang.
2022 — 2026
Tsinghua University
B.E. · Electronic Information Science and Technology
Department of Electronic Engineering. Thesis on reliable foundation-model transfer for industrial anomalous sound detection.
Related work05Publications & writing
Research, in writing.
Manuscripts, research software, and reports. Each work is listed with its current documented status.
Manuscript & research software
HVCR: Causal Evaluation of Large Multimodal Models in Human-like Video Reasoning
Repository-listed software citation
Co-authored work. No venue acceptance is asserted in the reviewed public record.
Bachelor’s thesis
Reliable Foundation-Model Transfer for Industrial Anomalous Sound Detection
Department of Electronic Engineering, Tsinghua University
The reviewed public summary is available on this site.
Course research report
SpecGuard: Channel-Aware Prompt Defense for Indirect Prompt Injection
Artificial Intelligence Safety and Governance course project
Results come from a recorded local benchmark, not an independently reproduced study.
06Recognition
Awards & recognition.
- 2019
Silver Medal · 20th Asian Physics Olympiad (APhO)
Adelaide, Australia