I study how intelligent systems are made to betray their purpose — across language models, multi-agent coordination, and behavioral biometrics. The work hunts for the structural condition that makes a whole class of failure inevitable.
▄▄▄▄▄▄▄ ▄▄▄▄▄▄▄ █ ▄▄▄ █ █▄ ▄▄▀█ █ ███ █ ▀▄▀▄▀ █ █▄▄▄▄▄█ █▀█ ▀▄█ ▄▄▄ ▄ ▄▄▄ ▄▀ ▄▄▄ ▀█▄▄ █▀▄▀▄▀ ▄▀ █▀▄▄▄▄█▀▀ ▄ ▄▀█ ▄▄▄▄▄▄▄ █▄▀█ ▀▀▄ █ ▄▄▄ █ ▀▀▀▄ ▀█ █ ███ █ ▀█▄ ▄▀▄▀ █▄▄▄▄▄█ ▄ ▄ ▀█▄
I work at the intersection of AI security, multi-agent systems, and applied cryptography — three domains sharing one structural question: what makes an entire category of system exploitable, not just a single weak instance?
That question produced a taxonomy of LLM adversarial surfaces (Best Paper, ICEM 2024), an ongoing MARL study of emergent coordination failure in drone swarms at DIAT (Defence Institute of Advanced Technology, a DRDO-affiliated deemed university), and a behavioral-biometric adversarial primitive (MIMIC, targeting IEEE TIFS). CTF results are execution evidence — not the headline.
A structural taxonomy of LLM adversarial attack surfaces — twelve categories across injection, jailbreaking, evasion, and extraction — designed as a map of failure conditions rather than a single exploit chain.
A QMIX-based multi-agent architecture for coordinated jammer avoidance in drone swarms. Phase D results: +13.0 detection points under chase-jammer via KL-anchored Navigator addressing emergent policy decoupling. Pre-publication.
An LSTM/DDPM hybrid synthesizing realistic human mouse-movement trajectories on a 212K+ sample dataset — full bot-detection evasion demonstrated against commercial detection systems.
How is an intelligent system made to betray its purpose — and what is the structural condition for that betrayal across language models, coordinating agents, and cryptographic trust protocols?
Genuine intellectual interest with no published output yet — bounded deliberately so it's never confused with verified research above.
QBER-threshold models in BB84 protocols may be structurally blind to source-layer compromise.
BB84/E91source compromisePQCModelling attack surface as a hypergraph; quantum-walk sampling as traversal for non-obvious pivots.
hypergraphsquantum walkspivot discoveryWhether quantum feature maps in variational classifiers offer genuine robustness — or relocate the attack surface.
VQCQML evasionNISQFast, local static analyzer combining Regex and AST to detect 30+ Python security vulnerabilities. Severity-graded findings across code execution, command injection, hardcoded secrets, weak crypto, and unsafe deserialization. Rich CLI with syntax-highlighted reports.
Python-based network scan detector for TCP Null/UDP scans with JARVIS-like voice alerts for critical ports (FTP, Telnet, SSH, etc.) and a real-time UI dashboard. Powered by Scapy for live packet capture and analysis.
A Python-based port scanning and service identification tool. Maps open ports to running services with banner grabbing and protocol fingerprinting. Designed for fast recon during offensive security assessments.
PowerShell-based automated threat-surface enumeration for Windows environments. Covers open ports, scheduled tasks, loaded modules, and startup extensions — generates a structured threat exposure report.
QMIX-based multi-agent RL for jammer avoidance. Phase D: +13.0 detection points via KL-anchored Navigator. Building toward IEEE TNNLS submission.
Secure infrastructure, VPN architecture, OWASP vulnerability remediation. Zero-trust implementation across production API surfaces.
Developed the ICEM adversarial taxonomy. Best Paper and Best Presentation at ICEM 2024 / NCTAAI 4.0.
National-ranked CTF competitor. Building RAV3N-SEC, SPH1NX, P.R.I.S.M, T.E.M.P.E.S.T. 4 public repositories, 33 stars.
Specializing in cybersecurity with active research publications, internships, and CTF circuit involvement alongside coursework.
Open to research collaboration, security consulting, CTF team invitations, and speaking — particularly in adversarial AI, multi-agent robustness, and applied cryptographic security.