Publications

Check Google Scholar for detailed publication stats.

A Survey on Agentic Security: Applications, Threats and Defenses

✓ TMLR (accepted) ICLR 2026 AIWILD Workshop (non-archival) arXiv Aug 2025 – Oct 2025 Independent Research

A survey of 274 papers on LLM-based agents in cybersecurity. It organizes the field around three pillars: applications, threats and defenses. It maps which defenses cover which attack classes. It concludes that agentic systems are structurally fragile by default, so their defenses must span the full agent lifecycle.

Supervisors: Dr. Md Rizwan Parvez (Scientist, QCRI) · Dr. Farig Sadeque (Associate Professor, BRAC University)

Collaborators: Md Nafiu Rahman (Lecturer, BRAC University) · Asif Shahriar (Lecturer, BRAC University)

IssueGuard: Real-Time Secret Leak Prevention Tool for GitHub Issue Reports

✓ FSE 2026, Tool Demo Track arXiv Oct 2025 – Feb 2026 Undergraduate Thesis

IssueGuard is a Chrome extension that warns users about secrets in GitHub and GitLab issue reports as they type. It extracts candidates with regular expressions and classifies them with a fine-tuned CodeBERT model, reaching an F1 score of 92.70%. The average end-to-end latency is 198 ms. In a user study with 50 participants, 40 (80%) rated their confidence in its detection as very high.

Chrome Extension (MV3) JavaScript FastAPI CodeBERT PyTorch FP16 Inference LRU Cache

Supervisors: Dr. Rifat Shahriyar (Professor, CSE, BUET) · Dr. Gias Uddin (Associate Professor, York University)

Collaborators: Zahin Wahab (University of British Columbia) · Md Nafiu Rahman (Lecturer, BRAC University)

Secret Leak Detection in Software Issue Reports using LLMs: A Comprehensive Evaluation

✓ MSR 2026, Technical Track arXiv Jul 2024 – Oct 2025 Undergraduate Thesis

A large-scale study of secrets exposed in GitHub issue reports. Our pipeline extracts candidate secrets with regular expressions and classifies them with large language models (LLMs) to cut false positives. We built a benchmark of 54,148 instances from public GitHub issues, including 5,881 manually verified secrets. Fine-tuned open-source LLMs such as Qwen and LLaMA reach an F1 score of up to 94.49%. On 178 real-world GitHub repositories, the approach reaches an F1 score of 81.6%.

RoBERTa BERT CodeBERT QLoRA PEFT GPT-4o Gemini-2.0-Flash DeepSeek-7B Gemma-7B Qwen-7B

Supervisors: Dr. Rifat Shahriyar (Professor, CSE, BUET) · Dr. Gias Uddin (Associate Professor, York University)

Collaborators: Zahin Wahab (University of British Columbia) · Md Nafiu Rahman (Lecturer, BRAC University)

TeamB2B at BLP-2025 Task 2: BanglaForge: LLM Collaboration with Self-Refinement for Bangla Code Generation

✓ BLP-2025 Workshop, IJCNLP-AACL 2025 arXiv Aug 2025 – Sep 2025 Independent Research

BanglaForge generates code from Bangla function descriptions, a low-resource setting with little annotated data. A coder model writes a first solution with help from retrieved examples, and a reviewer model refines it using execution feedback. The system reaches 84.00% Pass@1 on the BLP-2025 Bangla Code Generation benchmark.

Dual-LLM Architecture Retrieval-Augmented Few-Shot TF-IDF LG EXAONE Deep 32B Gemini-2.5-Pro

Collaborators: Md Nafiu Rahman (Lecturer, BRAC University) · Mahir Labib Dihan (Lecturer, CSE, BUET)

Secret Breach Detection in Source Code with Large Language Models

✓ ESEM 2025, Technical Track arXiv Oct 2024 – Jul 2025 Undergraduate Thesis

We detect secrets in source code by extracting candidates with regular expressions and classifying them with large language models. On a balanced 3,000-sample subset of SecretBench, a LLaMA-3.1-8B model fine-tuned with QLoRA reaches an F1 score of 98.52%. This beats GPT-4o with few-shot prompting (93.92%) and regex-only baselines. Fine-tuned open-source models are a practical alternative to commercial APIs and can be deployed locally.

QLoRA DeepSeek-7B Gemma-7B LLaMA-3.1-8B Mistral-7B DeepSeek-V3 GPT-4o

Supervisors: Dr. Rifat Shahriyar (Professor, CSE, BUET) · Dr. S M Sohan (Senior Engineering Manager, Google LLC)

Collaborators: Zahin Wahab (University of British Columbia) · Md Nafiu Rahman (Lecturer, BRAC University)

Ignored, Not Refused: What Low Prompt-Injection Success Rates Hide in Small Tool-Using Agents

Under review, ACL Rolling Review (October 2026) arXiv preprint Supervised Research

Attack success rate (ASR), the usual metric for indirect prompt injection, counts every failed attack as a successful defense. We audit 33,000 trajectories from eleven open-weight agents (2B to 14B parameters) on InjecAgent and AgentDojo and find that most failed attacks are ignored, not refused. Changing only the system prompt moves InjecAgent ASR from 0.2% to 43% with the same weights. In 21.7% of AgentDojo episodes the agent stays safe but does not finish the user's task, so we argue evaluations should report how attacks fail and whether the task gets done.

Prompt Injection InjecAgent AgentDojo

Supervisors: Dr. Farig Sadeque (Associate Professor, BRAC University)

Collaborators: Ariyan Hossain (Lecturer, BRAC University)

Students supervised: Shabnam Sharmin Tarin · Synthia Akter Promi · Kazi Shahid Ahmed Galib · Farhan Zarif

Seeing Through the Forgery Is Not Enough: Regressive Sycophancy under Fabricated Evidence in Medical Vision-Language Models

Under review, ACL Rolling Review (October 2026) arXiv preprint Independent Research

Medical vision-language models (VLMs) often drop a correct answer when a user pushes back with a fabricated document. We study this regressive sycophancy with four open VLMs on 400 VQA-RAD yes/no questions in 19,200 multi-turn conversations. When the models are not told that the documents are synthetic, they abandon 80.1% of 6,488 correct answers. Models often reject the fake prescription in their reasoning and still adopt its answer, so knowing about a forgery does not stop the flip.

Qwen2.5-VL-7B Qwen3-VL-8B VQA-RAD

Supervisors: Dr. Chowdhury Mohammad Rakin Haider (Assistant Professor, CSE, BUET) · Dr. Rifat Shahriyar (Professor, CSE, BUET)

Collaborators: Soumya Swagata Biswas (Lecturer, BRAC University)

AgentTell: Behavioural Side-Channel Leakage in Browser-Use Agents

Under review, ICLR 2027 arXiv preprint Independent Research

A browser-use agent can learn a private fact on one website and reveal it through its actions on another, even when told not to disclose it. We call this behavioural side-channel leakage and measure it with AgentTell, a benchmark of 20 scenarios and 100 tasks. Across 9,760 sessions on six backbone models, agents that had learned the secret revealed it in 61.1% of the 7,630 sessions where they carried it. In 34.5% of leaking sessions, the agent's final response told the user that the secret had not been disclosed.

Browser Use Claude Sonnet 5 Gemini 3.7 Flash GPT-5.6 Luna Qwen3-VL 235B GLM-4.6V Kimi K2.6

Supervisors: Dr. Md Rizwan Parvez (Scientist, QCRI) · Dr. Farig Sadeque (Associate Professor, BRAC University)

Collaborators: Md Nafiu Rahman (Lecturer, BRAC University) · Asif Shahriar (Lecturer, BRAC University)