|
Jafar Isbarov (Cəfər İsbarov) I'm a second-year Ph.D. student at the Data Security and Privacy Lab, Virginia Tech. I am working on AI security under the guidance of Dr. Murat Kantarcioglu. Before that, I was a visiting student at New York University where I was working with Dr. Duygu Ataman on evaluation of multilingual LLMs. I also have 3+ years of industry experience building NLP systems in low-resource environments. My current research interests include:
In the past, I have worked on various problems, such as (1) single-cell data analysis, (2) spelling correction for agglutinative languages, and (3) foundation models for Azerbaijani. External: Email / CV / Scholar / GitHub / Hugging Face / LinkedIn Go to: Writing / Publications / Education / Academic Service / News |
|
|
Writing |
|
Jul '26 |
We cannot simulate AI security research, LessWrong. |
|
Publications |
|
Jafar Isbarov, Ulvi Suleymanov, Ilia Shumailov, Murat Kantarcioglu arXiv, 2026 We present GitInject, an open-source framework for evaluating prompt injection attacks in real, live GitHub Actions workflows. Across four AI providers, every one is vulnerable to at least one attack class in its default configuration. |
|
K Venkatesh, Jafar Isbarov, S Amin, Murat Kantarcioglu, Jiaming Cui arXiv, 2026 Code We propose CASPIAN, a framework for online detection and attribution of cascade attacks in LLM multi-agent systems, using cross-channel causal monitoring to trace how adversarial influence propagates between agents. |
|
Jafar Isbarov, Murat Kantarcioglu arXiv, 2026 Code We show that monitoring-based defenses can be systematically bypassed via Agent-as-a-Proxy attacks. |
|
Jafar Isbarov, Arofat Akhundjanova, Mammad Hajili, Kavsar Huseynova, Dmitry Gaynullin, Anar Rzayev, Osman Tursun, Ilshat Saetov, Rinat Kharisov, Saule Belginova, Ariana Kenbayeva, Amina Alisheva, Aizirek Turdubaeva, Abdullatif Köksal, Samir Rustamov, Duygu Ataman ACL 2025 (Main Conference) Code / Dataset We introduce TUMLU, a unified and native language understanding benchmark for Azerbaijani, Crimean Tatar, Karakalpak, Kazakh, Tatar, Turkish, Uyghur, and Uzbek. This paper was accepted to the main conference at ACL 2025 |
|
Jafar Isbarov, Kavsar Huseynova, Elvin Mammadov, Mammad Hajili, Duygu Ataman SIGTURK Workshop, ACL, 2024 Dataset We introduce aLLMA, a series of foundation models for Azerbaijani language, alongside the pretraining corpus and an NLU benchmark. |
See my Google Scholar page for the full list.
|
Education |
|
PhD in Computer Science, Virginia Tech
08.2025 - Present GPA: - Advisor: Dr. Murat Kantarcioglu |
|
MSc in Computer Science, George Washington University
08.2023 - 12.2024 GPA: 3.92 / 4.00 Advisors: Dr. Duygu Ataman (New York University) & Dr. Samir Rustamov |
|
BSc in Biomedical Engineering, Azerbaijan State Oil and Industry University
09.2018 - 06.2023 GPA: 3.40 / 4.00 |
|
Academic Service Reviewer at:
Officer at ACL Special Interest Group in Turkic Languages (2025-2026). |
|
News |
|
Jun. '26 |
New paper: GitInject, on real-world prompt injection attacks in AI-powered CI/CD pipelines. |
|
May '26 |
New paper: CASPIAN, on detecting and attributing cascade attacks in LLM multi-agent systems. |
|
Feb. '26 |
New paper: Bypassing AI Control Protocols via Agent-as-a-Proxy Attacks. |
|
Aug. '25 |
Started Ph.D. program in Computer Science at Virginia Tech. |
|
May '25 |
Our paper TUMLU has been accepted to ACL 2025 (Main Conference)! |
|
Dec. '24 |
Graduated from George Washington University with a Master's degree in Computer Science. |
|
Jun. '24 |
Our paper Open Foundation Models for Azerbaijani Language has been accepted to SIGTURK Workshop, ACL 2024! |
|
The source for this website is from here. |
|
|