Pritom Bhowmik

AI Security Researcher/Trustworthy AI Agents & Adversarial ML

Research Area: Trustworthy AI AgentsAdversarial Machine LearningMemory & Retrieval PoisoningSecurity–Utility Cost of DefensesMulti-Agent TrustHuman–AI Interaction

Portrait of Pritom Bhowmik
01

Research

I work on trustworthy AI agents and adversarial machine learning. My current research asks what it costs to defend an agent’s memory against poisoning, and how agents in multi-agent systems can trust the memory they share. I am also interested in human–AI interaction.

02

Publications

  1. 2026
    The Price of Safety: Benign-Case Utility and Token Overhead of Memory-Poisoning Defenses in LLM Agents
    Pritom Bhowmik
    arXiv:2609.22818 [cs.CR], 2026. Submitted, Workshop on Agentic Engineering (AGENT) at ICSE 2027.
    [PDF][arXiv]
  2. 2026
    Outcome-Bound Trust: Falsifiable Claims as a Security Primitive for Memory in Agent-to-Agent AI Systems
    Pritom Bhowmik and Hao Liu
    To be submitted, Workshop on Trustworthy Evolution of AI Agents at AAAI 2027.
  3. 2022
    Machine Learning in Production: From Experimented ML Model to System
    Pritom Bhowmik
    ScienceOpen Preprints, 2022.
    [PDF][DOI]
  4. 2021
    Vulnerability of Neural Networks to Adversarial Attack and Defenses in Computer Vision
    Pritom Bhowmik
    IEEE Computer Society, Bangladesh Chapter, 2021. Abstract.
  5. 2019
    Research Study on Basic Understanding of Artificial Neural Networks
    Pritom Bhowmik
    Global Journal of Computer Science and Technology: D Neural & Artificial Intelligence, 19(4), pp. 5–7, 2019.
    [PDF][Paper]
03

Research & Teaching Experience

Research

  • Researcher, AI SecuritySep 2025 – Present
    Memory Security in LLM Agent Systems
    What defending agent memory costs, and how agents can trust memory they share.
  • Graduate Researcher, Data Science LabJan 2024 – May 2025
    Montclair State University · Advisor: Dr. Hao Liu
    Multi-agent RAG for pharmaceutical supply-chain decisions: query routing across retrieval, analytics, and web-search agents; retrieval accuracy 72% → 97%, latency 3.2 min → 12 s.

Teaching

  • Adjunct Instructor, Computer ScienceSep 2025 – Jul 2026
    Monroe University, New Rochelle, NY
    Programming Logic, Python, Database Systems, and Data Science (undergraduate and graduate). Designed Python, C++, and SQL labs for 100+ students per semester.
04

Current Research

2026 · Submitted, AGENT @ ICSE 2027

The Price of Safety

  • Measures what memory-poisoning defenses cost an LLM agent on attack-free traffic, on the LoCoMo benchmark.
  • Write-time defenses: no detectable utility cost.
  • Read-time reranker: −4.4 points accuracy; wrongly flags 9.9% of retrieved memories, removing memory from 45% of answers.

[PDF][arXiv]

With Hao Liu · To be submitted, AAAI 2027 workshop

Outcome-Bound Trust

  • Turns agent messages into checkable claims, and caps what an agent commits to each counterparty by the claims it has already honored.
  • In a supply-chain simulation: cut losses from lying agents by 98.1% at a 5.43% utility cost.
  • Safety invariants model-checked in TLA+.
05

Education

  • M.S., Data ScienceMay 2025
    Montclair State University

    GPA 3.97 / 4.00
    Alpha Epsilon Lambda Honor Society

  • B.Tech., Computer ScienceJul 2020
    Institute of Engineering and Management, Kolkata