V.MOTHUKURI

VIRAAJI MOTHUKURI

> AI RESEARCHER // PHD CANDIDATE // EX-VP DATA SCIENCE @ JPMC

[SCHOLAR] [GITHUB] [LINKEDIN]

// 01 Research

Research Focus

Causal foundations, security, and safety of AI systems.

01

LLM Security, Alignment & Mechanistic Interpretability

Building mechanistic interpretability methods for detecting deception, sandbagging, and alignment faking in large language models. Sparse autoencoder probes, activation-level steering, and counterfactual evaluation on open-weight models. Forthcoming work at ACL 2026 and ICML 2026. Companion line on adversarial robustness, jailbreak resilience, and verification of safety training under distribution shift.

02

Agentic AI Security

Analyzing security threats in autonomous AI agents that plan, use tools, and interact with external systems. Research spans vulnerability discovery in multi-agent coordination, attack surface analysis for tool-using agents, and causal approaches to detecting adversarial manipulation of agent behavior.

03

Causal AI & Agent Safety Verification

Applying causal inference to verify and audit AI agent behavior. Developing methods for causal structure discovery in learned representations, temporal causal analysis of black-box agents, and identifiability frameworks that bridge causal reasoning with deployed AI systems.

04

Post-Quantum Cryptography & AI

Researching the intersection of quantum computing threats and AI-driven security solutions. Focus on AI-powered cryptographic inventory discovery, risk assessment for quantum-vulnerable systems, and automated migration strategies to NIST-standardized post-quantum algorithms.

// 02 About

About

I am a PhD candidate in Computer Science at Kennesaw State University (defense December 2026). My dissertation builds mechanistic interpretability tools for detecting deception, sandbagging, and alignment faking in large language models, using sparse autoencoder probes and activation-level steering on open-weight models.

Before the PhD I was Vice President and Lead Data Scientist at JPMorgan Chase, where I co-invented a patented trade-surveillance ML system (US Patent 12,536,582) that shipped into a regulated production environment and won two industry awards. That work taught me what production constraints, compliance review, and adversarial settings actually do to research methods, and it shapes the work I do now.

My current research extends interpretability methods to agentic AI systems and post-quantum cryptography risk assessment, including "Causal Detection of Multi-Step LLM Agent Attacks" (ICML 2026) and "Trajectory Signatures of Deception in Large Language Models" (ACL 2026). Three Best Paper Awards (IEEE IoT Journal 2025, IEEE Blockchain 2024, FGCS 2022).

// 03 Safety & Alignment

// ADVISORY: AI SAFETY, ALIGNMENT & SECURITY

AI Safety & Alignment Research

[DECEPTIVE ALIGNMENT] [CAUSAL VERIFICATION]

Alignment Verification & Deceptive Behavior

  • Developing methods to identify when AI systems exhibit strategic deception or alignment faking behaviors
  • Applying causal inference to verify whether safety training produces genuine behavioral change or shallow compliance
  • Creating frameworks to test AI behavior consistency across different contexts and prompting strategies
  • Investigating alignment faking as a capability-dependent phenomenon in large language models

[AGENTIC SECURITY] [ADVERSARIAL ROBUSTNESS]

Agentic AI & LLM Security

  • Analyzing security vulnerabilities in autonomous AI agents that use tools, access data, and interact with external systems
  • Developing robust defenses against prompt injection and jailbreak attacks in LLMs
  • Temporal causal discovery methods for extracting objectives and predicting vulnerabilities in black-box agents
  • Ensuring integrity and security of AI model pipelines from training to deployment

AI SECURITY RESEARCH HUB

Access our collection of AI security research, vulnerability assessments, and defensive techniques.

EXPLORE PLATFORM

KEY RESEARCH CONTRIBUTIONS

  • Developed mechanistic interpretability probes for alignment-faking detection on open-weight LLMs; preliminary results show separability of "model-knows-it-is-being-evaluated" features from task-encoding features (ACL 2026, ICML 2026).
  • Authored LLMSmartSec, an LLM-driven smart-contract auditing framework using annotated control flow graphs (IEEE Blockchain 2024). Extended to LLM-as-judge evaluation (ICBC 2025).
  • Co-invented and shipped a patented trade-surveillance ML system at JPMorgan Chase (US Patent 12,536,582). Two industry awards.
  • Wrote the field-defining federated learning security survey (Future Generation Computer Systems 2021; 1,928 citations as of 2026).
  • Reviewer for ICML, NeurIPS, and multiple IEEE and Elsevier venues; IEEE Senior Member.

// 04 Experience

Professional Experience

AUG 2023 — PRESENT Kennesaw State University

Research Assistant (Doctoral Candidate)

Leading research on AI security, causal inference for agent safety, and LLM alignment. Developing causal structure discovery methods for verifying learned agent representations and creating automated security testing frameworks for autonomous agents.

JUL 2021 — AUG 2023 JP Morgan Chase

Data Scientist Lead, Vice President

Led award-winning and patented work on trade surveillance. Integrated news, market, and trade data to identify suspicious trading activity. Architected ML pipelines on AWS cloud with MLOps implementation and applied NLP for insider trading detection.

AUG 2019 — JUL 2021 Kennesaw State University

Research Assistant

Conducted research on Federated Learning, Blockchain integration, and ML model quantization. Published multiple papers on security and privacy of federated learning, worked with frameworks like PySyft and TensorFlow Federated.

OCT 2016 — AUG 2019 JP Morgan Chase

Data Specialist

Supported the Emerging Payments division and ChasePay app. Managed lifecycle and reconciliation of user data across multiple databases, automated mundane tasks, and developed innovative strategies for knowledge transfer.

// 05 Publications

Publications

Selected work organized by publication year.

[1] V Mothukuri et al. (2026)

"Trajectory Signatures of Deception in Large Language Models"

Annual Meeting of the Association for Computational Linguistics (ACL)

[2] V Mothukuri et al. (2026)

"Causal Detection of Multi-Step LLM Agent Attacks"

International Conference on Machine Learning (ICML)

[5] V Mothukuri (2025)

"GRP-071 Next-Generation DAPPs Development with Self-Service AI Agents"

2025

[6] V Mothukuri et al. (2025)

"AgentFL: AI-Orchestrated Agents for Federated Learning"

IEEE International Conference on Distributed Computing Systems (ICDCS)

[7a] RM Parizi, V Mothukuri (2025)

"System and Method for Security and Trustworthiness of Agentic AI"

US Patent, Docket 065848.0052 (to appear)

US PATENT — TO APPEAR
[7b] V Mothukuri et al. (2026)

"Method and System for Performing Compliance Reviews"

US Patent Application

US PATENT — PENDING
[10] V Mothukuri, RM Parizi, S Pouriyeh, A Mashhadi (2022)

"Cloudfl: a zero-touch federated learning framework for privacy-aware sensor cloud"

17th International Conference on Availability, Reliability and Security

CITED: 4
[17] V Mothukuri (2021)

"Federated Learning for Secure Sensor Cloud"

2021

// 06 Skills

System Specs

AI & ML Causal Inference, AI Safety & Alignment, LLM Security, Agentic AI, Federated Learning, Blockchain Security
Programming Python, Java, Go, C/Pro C, Shell Scripting, JavaScript
Frameworks PyTorch, TensorFlow, Keras, scikit-learn, Docker, Kubernetes
Cloud & Security AWS, Google Cloud, Post-Quantum Cryptography, Hyperledger Fabric, Cybersecurity, IoT Security

// 07 Awards

Awards & Recognition

Viraaji Mothukuri receiving an award
[2025] IEEE Internet of Things Journal Best Paper Award
[2024] IEEE Blockchain Best Paper Award
[2022] FGCS Best Paper Award
[JPMC] American Financial Technology Award — Best Compliance Initiative
[JPMC] Fintech Futures Banking Tech Award — Best Use of RegTech
[KSU] Best PhD Student — Kennesaw State University
[JPMC] Shining Star Award — JPMorgan Chase