All ProjectsContact Dr. Hossain
LLM SecurityAgentic AIAdversarial ML
Agentic LLM Security
As large language models are deployed as autonomous agents that invoke tools, browse the web, and coordinate with other agents, they introduce a new class of attack surfaces spanning prompt injection, tool abuse, memory poisoning, and multi-agent collusion. This project develops threat models, detection mechanisms, and runtime defenses that make agentic LLM systems robust, auditable, and trustworthy in high-stakes deployments.
Research Objectives
- Characterize attack surfaces unique to tool-using, memory-augmented LLM agents
- Build runtime monitors that detect prompt injection and anomalous tool invocations
- Design policy-enforcing guardrails for multi-agent communication and task delegation
- Develop benchmarks and red-teaming frameworks for evaluating agentic LLM robustness
Methods & Techniques
- Adversarial prompt generation and automated red-teaming pipelines
- Information-flow tracking across agent memory, tools, and external content
- Constrained decoding and policy-based action filtering for safe tool use
- Reinforcement learning from safety feedback (RLSF) for defensive agent training
Interested in this research?
Get in touch to discuss collaboration or graduate opportunities.