Research
Why agents fail, and how to measure it.
Most of my engineering work is building agents that run in production. My research looks at the same systems from the other side: where they break, what an attacker can do with them, and whether training changes them in ways nobody asked for.
I work on three things: how fine-tuning can shift a model's behaviour outside its target domain, the security of agents that can act in real environments, and data for testing agents in Bangla.
- Program
- MSc in Computer Science & Engineering, RUET
- Supervisor
- Dr. Md. Rabiul Islam
- Since
- Nov 2025
- Focus
- Agent security · LLM alignment · Evaluation · Agentic datasets
Three lines of research.
Ongoing
MSc thesis, RUET
Alignment drift after fine-tuning
How narrow fine-tuning can shift a language model's behaviour well outside the task it was trained on (emergent misalignment), and which mitigations work. The work starts by reproducing a published result on a controlled pipeline before testing anything new.
Alignment · Emergent Misalignment · Fine-tuning · Evaluation
arXiv preprint
2026
Cyber-Capable AI Agents: Vulnerabilities, Evaluation Containment, and Defensive Response
A structured review of what goes wrong when cyber-capable models are given autonomy: multi-step attack chains, privilege separation, provenance, and who can reach a responder when an evaluation escapes its sandbox.
Agent Security · Evaluation Containment · AI Safety
Read the preprint →Ongoing
Dataset
Jontro: agentic data for Bangla
A tool-use dataset for Bangla-language agents. Tool results come from deterministic simulators, so every trajectory can be replayed and checked, and a first-call diagnostic tests whether a model picks the right tool in Bangla at all.
Agents · Bangla · Datasets · Tool Use
Essays on agent safety.
Can We Really Make AI Safe?
Injection Is Not Influence: The Illusion of LLM Memory
Earlier publications.
Applied machine learning papers from before the shift to agents.
2024 · arXiv Preprint
Comparative Performance of Machine Learning Algorithms for Early Genetic Disorder and Subclass Classification
Abu Bakar Siddik, Faisal R. Badal, Afroza Islam
Evaluates machine-learning models for early genetic-disorder class and subtype prediction from clinical indicators available at birth or infancy.
2023 · Research Publication
Unraveling the Enigmatic Frontier: Deciphering the Distinction Between AI-Generated and Real Images
Abu Bakar Siddik et al.
Investigates the boundary between AI-generated and authentic images using deep learning classification techniques, with implications for digital forensics and media authenticity.
2023 · Research Publication
Real-time Patient Monitoring System to Reduce Medical Error with the help of Database System
Abu Bakar Siddik et al.
Proposes a real-time patient monitoring architecture using database-backed alerts to reduce clinical errors and improve healthcare outcomes.