I research the security of machine-learning and agentic AI systems — how they fail under adversarial pressure, and how to detect and mitigate those failures. My background is in offensive and purple-team security; I work where that experience meets rigorous ML evaluation.
Master's Cybersecurity · UPES Dehradun · Class of 2027
MIVL — Memory Integrity Verification Layer A defense layer for detecting and mitigating memory-poisoning attacks on agentic AI systems. Defines a set of detection signals over agent memory writes, a scoring function, and a remediation pipeline, evaluated on a labelled dataset spanning four attack classes — direct memory injection, accumulation, retrieval manipulation, and cross-agent propagation — built from real prompt-injection data and synthetic novel classes. Paper and fully reproducible pipeline included.
Silent Exposure — Metadata Leakage in Encrypted Messaging A comparative study of metadata leakage across Signal, WhatsApp, Telegram, iMessage, Messenger, and RCS, and how that leakage enables adversarial persona reconstruction. Reproducible end-to-end: synthetic data generation, statistical analysis, figures, and paper. All data is synthetically generated from published leakage parameters — no real users monitored, no live traffic captured.
- Extending MIVL's threat model and stress-testing its detection signals against adaptive attackers.
- Reading toward red-team methodology for LLM-based and agentic systems.