Chirag Dewan
AI security researcher — red teaming LLMs & agents.
Offensive security background.
Selected work
MCP-Poison-Bench
A reproducible benchmark and client-side defense for MCP tool-poisoning.
Built the defense, measured it (ASR 1.00 → 0.00), then broke it (9 / 11 techniques bypass).
PARALLAX
Behavioral threat detection for AI platforms — metadata-only, no payload inspection.
0.68 AUC on 17M real authentication events.