AI For Security Research Areas

LLM Jailbreak Attacks & Defenses

Developing systematic frameworks for evaluating and hardening LLMs against jailbreak attacks. Covers both attack generation (multi-agent RL-based obfuscation) and defense strategies, with emphasis on black-box and white-box threat models. Current work includes cipher-based attacks, multi-agent coordination patterns, and time-persistent jailbreak methodologies.

Paper 1: MetaCipher: A Time-Persistent and Universal Multi-Agent Framework for Cipher-Based Jailbreak Attacks for LLMs

Vulnerability Detection & Analysis

Offensive Security Benchmarks

VLM Hallucination Mitigation

Efficient AI for Code Generation

AI-driven vulnerability detection and analysis for C/C++ code and binary programs. Combines retrieval-augmented generation (RAG) with multi-agent LLM pipelines for memory corruption analysis. Current work focuses on scalable detection across large codebases and fine-tuning specialized models for vulnerability reporting.

Paper 1: RAVEN: Retrieval-Augmented Vulnerability Exploration Network for Memory Corruption Analysis in User Code and Binary Programs

Building scalable, open-source benchmark datasets for evaluating LLM capabilities in offensive security tasks. Covers Capture The Flag (CTF) challenges, penetration testing scenarios, and vulnerability exploitation. Focused on creating reproducible evaluation standards for the security community.

Paper 1: NYU CTF Bench: A Scalable Open-Source Benchmark Dataset for Evaluating LLMs in Offensive Security

Paper 2: An Empirical Evaluation of LLMs for Solving Offensive Security Challenges

Grounding vision-language models (VLMs) with object detection to mitigate counting hallucinations. Develops modular architectures that combine traditional computer vision detectors with VLMs for reliable visual reasoning. Aims to make VLMs trustworthy for security-critical visual analysis tasks.

Paper 1: GroundCount: Grounding Vision-Language Models with Object Detection for Mitigating Counting Hallucinations

Cost-efficient and reliable code generation using cascaded multi-model frameworks with self-testing. Reduces inference costs while maintaining code quality through intelligent model routing and verification. Applicable to both general code completion and security-sensitive code generation tasks.

Paper 1: Model Cascading for Code: A Cascaded Black-Box Multi-Model Framework for Cost-Efficient Code Completion with Self-Testing

© 2026 eBRAIN