Adversarial Systems · India · 2026 Research intern at DIAT (DRDO). Best Paper, ICEM 2024. I break intelligent systems to learn why whole classes of them can be broken.
↓ Scroll

Systems betray their purpose. I find where.

Samratth Singh — adversarial ML, multi-agent robustness, and behavioral threat modeling. The work hunts for the structural condition that makes a whole class of failure inevitable, not a single clever exploit.

Open to research collaboration

Selected work

Research · Systems
01
LLM Attack Taxonomy
★ Best Paper · Best Presentation — ICEM 2024 / NCTAAI 4.0

A structural taxonomy of LLM adversarial surfaces — twelve categories spanning injection, jailbreaking, evasion, and extraction. A map of failure conditions, not an exploit chain.

Published
02
SWARN — MARL Swarms

QMIX-based multi-agent RL for drone-swarm jammer avoidance at DIAT (DRDO). Identified emergent policy decoupling and traced it to mechanism via a KL-anchored Navigator. Submitted to IEEE TNNLS.

In review
03
MIMIC — Motion Synthesis

Hybrid LSTM/Diffusion synthesis of human mouse trajectories, studied as a behavioral-biometric adversarial primitive. Probes the limits of motion-based bot detection. Target: IEEE TIFS.

In progress
04
PhishByte

A custom PyTorch MLP phishing detector engineered from scratch — F1 0.9484 — deployed on Hugging Face and past 60 downloads with zero marketing. Detection as a shipped artifact, not a notebook.

60+ downloads
05
T.E.M.P.E.S.T

Automated threat modeling and attack-surface enumeration for LLM-integrated enterprise deployments. Maps adversarial pathways, generates attack trees, quantifies exposure across boundaries.

Building

About

Profile

I work at the intersection of AI security, multi-agent systems, and applied cryptography — three domains sharing one question: what makes an entire category of system exploitable?

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 (DRDO), and a behavioral-biometric adversarial primitive. CTF results are execution evidence — not the headline.

Final-year Computer Science (Cybersecurity) at DY Patil International University, Pune. Chair of the ACM Student Chapter. Currently building toward IEEE Q1 publication and international graduate study.

AffiliationDIAT · DRDO
HonorsICEM 2024 · NCTAAI 4.0
BasedPune, Maharashtra
StatusOpen to collaboration

Execution evidence

Cyber range · Credentials
14th
National rank · Ciphathon CTF
5%
Top — HackTheBox
7%
Top — TryHackMe
12+
Attack categories mapped
CEH v12 — EC-Council Advanced Network Attacks Web Application Hacking Cryptography Fundamentals HTB Pro Hacker

Open notebook

Research in motion — pre-publication
Active — project EVELYN

Hypergraph threat recon

Modelling attack surface as a hypergraph, with quantum-walk sampling as a traversal method for non-obvious pivots — now being developed as EVELYN, applying quantum-walk graph ML to phishing infrastructure fingerprinting.

Log

Experience
2025 — Present

Research Intern — MARL for drone-swarm coordination

DIAT (DRDO-affiliated)

QMIX-based multi-agent RL for jammer avoidance. Identified emergent policy decoupling through behavioral observation and traced it to mechanism. Building toward IEEE Q1 submission.

2025 · 3 mos

Software Engineer Intern

Dexpert Systems

Secure infrastructure, VPN architecture, OWASP vulnerability remediation, and zero-trust implementation across production API surfaces.

2024

Research Intern — AI security

ICAR

Developed the ICEM taxonomy of LLM adversarial surfaces. Best Paper and Best Presentation at ICEM 2024 and NCTAAI 4.0.

2023 — Ongoing

Independent security researcher

Self-directed · CTF

National-ranked CTF competitor across Ciphathon, HackTheBox, TryHackMe, and SecLeaf. Builder of RAV3N-SEC and T.E.M.P.E.S.T.

Let's break
something together.