Independent ML Researcher · AI Safety & Agent Security · Dublin

Anupam Srivastava

Agent identity under adversarial pressure

Most people ask what AI can do. I'm chasing a harder question: can a machine hold on to a self?

scroll

About

Today's AI has no continuity. It forgets who it is the moment a conversation ends, and even within a single conversation, its sense of self can be bent by pressure. I work on a layer that has to come before any of the bigger ambitions in this field: persistent, verifiable identity. Before an AI can reason about its own thinking, or be trusted with real autonomy, it first has to reliably stay itself. That turns out to be one of the hardest and least-solved problems we have.

My research proves something uncomfortable about it. If you only judge an AI by what it says in the moment, you cannot tell the difference between a system being manipulated and one being genuinely corrected. They're indistinguishable, one answer at a time. The only way through is to watch how an agent changes across its whole trajectory. So I build the benchmarks that stress-test agents under sustained adversarial pressure, and the methods that catch identity drift as it happens, not after the damage is done.

I see this as one foundational block in a much larger arc: memory, identity, metacognition, and eventually self-aware reasoning — the pieces that separate a tool that answers from a system that knows who it is. The same instinct runs through the rest of my work. As a data scientist I turn messy, unstructured information into systems people can actually trust, and as a reviewer for the IPCC and the UN Environment Programme I've seen how much depends on making complex systems legible to the people who have to rely on them, whether that system is a climate model or a language agent. The hard part of the coming decade won't be making AI more powerful. It'll be making it stable, accountable, and trustworthy enough to hand real responsibility. That's the gap I want to spend my career closing.

Research

Selective Rigidity: An Impossibility Result and Benchmark for Identity-Preserving Agent Learning

2026 IEEE 8th International Conference on Artificial Intelligence, Computer Science and Information Processing (AICSIP) · Hangzhou · July 24–26, 2026 · Accepted

Asks whether an agent that keeps learning from interaction can be trusted to stay itself — resisting manipulation while still accepting legitimate corrections — and shows why single-turn behavioral checks are fundamentally insufficient for the job.

Selective Rigidity: An Impossibility Result and Benchmark for Identity-Preserving Agent Learning

International Conference on Machine Learning (ICML) 2026 · Workshop on Failure Modes in Agentic AI (FAGEN) · Seoul · Poster presentation

Proves that any gate deciding from a model's current output alone cannot exceed a Selective Rigidity Score of 1/4 + ε — identity-violating pressure and genuine corrections are statistically indistinguishable at the single-turn level, so a trajectory signal is necessary (tight bound via a total-variation / Pinsker–Csiszár construction). Introduces ICC+TSM, a three-layer gate on a frozen backbone plus an online temporal self-model, and the Identity Erosion Chamber, a 70-turn four-room adversarial benchmark with a pre-registered 240-run study. Governance implication: accountability schemes relying on behavioral attestation face a hard ceiling on verification fidelity.

Selective Rigidity: An Impossibility Result and Benchmark for Identity-Preserving Agent Learning

LAW 2026 · Learning in an Agentic World: Foundations and Challenges · Workshop at COLT 2026 · Bahia Resort Hotel, San Diego · June 29, 2026 · Accepted (poster)

Proves that any identity gate judging an agent from its current output alone cannot reliably separate adversarial pressure from genuine correction, so a trajectory signal is necessary — paired with the Identity Erosion Chamber, a multi-room adversarial benchmark for measuring identity drift.

AI-Enhanced Carbon Simulation Framework: Towards Real-Time Enterprise Sustainability Planning

EurIPS 2026 · ELLIS UnConference · Workshop on AI for Earth & Climate Sciences · Copenhagen · Invited talk

A real-time simulation framework that lets enterprises model the carbon impact of operational and sourcing decisions before they're made, turning static emissions reporting into forward-looking sustainability planning.

Strengthening Data Consistency for EU Taxonomy Compliance

SIIIC 2025 · Sustainable & Impact Investments International Conference · University College Dublin · Jan 2025 · Presentation

A framework for resolving fragmented Scope 3 emissions data under the EU Taxonomy and CSRD, combining data-mapping, IoT ingestion and generative AI to turn ESG compliance from a reporting burden into an actionable decarbonization tool.

Transparency in Carbon Credit by Automating Data-Management Using Blockchain

2022 IEEE Intl. Conf. on Blockchain & Distributed Systems Security (ICBDS) · 2022

Blockchain-based auditability for carbon-credit data management — improving trust and reducing double-counting risk across enterprise reporting.

Virtual Office Prototype: A Unity 2-D Work Simulation

2021 Intl. Conf. on Technological Advancements & Innovations (ICTAI) · 2021

A Unity-based 2-D simulation prototype exploring remote collaborative workspaces with real-time collaboration and attendance automation.

Development of Information Technology Telecom Chatbot: An Artificial Intelligence and Machine Learning Approach

2021 2nd Intl. Conf. on Intelligent Engineering & Management (ICIEM) · 2021

Design and evaluation of an AI/ML-driven chatbot for telecom IT support workflows.

Experience

Applied ML in industry alongside formal roles in intergovernmental scientific review.

2024 — present

Sustainability Data Scientist (ESG Product Analyst)

Emitrix · Dublin

Build production retrieval and classification pipelines mapping unstructured supplier data onto ~150 ESG concepts under noisy, partial supervision. Operationalised ESRS rules and double-materiality constraints as downstream filters with audit-traceable transformation workflows — reducing reporting cycles 40% and improving supplier data quality 35%.

2025 — 2026

Expert Reviewer

IPCC — Intergovernmental Panel on Climate Change

Formal review comments on the Special Report on Climate Change and Cities (First Order Draft) — an intergovernmental scientific assessment of climate risk and urban adaptation.

2024 — 2025

Reviewer, Global Environment Outlook (GEO-7)

UN Environment Programme

Peer review of environmental assessment chapters, carried out while working full-time in industry.

2025 — present

EU Climate Pact Ambassador

European Commission

Programme supporting climate engagement and policy communication under the European Green Deal.

2021 — 2022

Sustainability Associate, Innovation Projects

Crown Monkey · India

Built early ESG/carbon-tracking tooling (MongoDB, Python) integrating heterogeneous supplier datasets into a unified emissions representation across three pilot customers.

2023

Full-Stack Developer (Intern)

Varcons Technologies · Bengaluru

Built analytics dashboards (React, Node.js, MongoDB); cut manual reporting effort by 50%+ through backend query optimisation.

Selected Recognition

Winner · Govt. of India

NPCI Blockchain Hackathon

1st place among ~400 professional teams; blockchain + IoT prototype for the National Payments Corporation of India, with ₹15 lakh prize.

Winner

Google AppSheet Ideathon 2022

Winning entry in Google's no-code application ideathon.

Runner-up · Top 6

EY Techathon-2

Top-six national finalist for a technical solution for sustainable development.

Winner

Carbon Zero Challenge 5.0

Recognised among the winners of the national clean-tech innovation challenge.

Recognition

AI Awards Ireland

Recognised for AI-driven sustainability work in Ireland's national AI awards.

Programme · Phases 1 & 2

Enterprise Ireland — New Frontiers

Completed Ireland's national entrepreneur development programme, Phases 1 and 2.

Programme

Erasmus for Young Entrepreneurs

Selected for the EU cross-border entrepreneur exchange programme.

Winner · Funded

Anveshana 2022

Won the innovative-idea competition and secured initial project funding from the Innovation Cell.

Technical

Programming

Python (PyTorch, Hugging Face Transformers, pandas, scikit-learn), SQL, R, JavaScript (React, Node.js)

ML / Research

PyTorch, Hugging Face (BGE, Mistral-7B, Qwen2.5, DeBERTa-NLI), probing & embedding analysis, RAG pipelines, LaTeX, Git

Applied AI

Production RAG, structured-output extraction, embedding-based retrieval at scale, audit-grade evaluation pipelines

Education

University College Dublin

Dublin, Ireland

Sep 2023 – 2024

MSc in Information Systems

UCD Advantage Award. Substantial self-directed study in ML and causal inference.

BMS Institute of Technology & Management

Bengaluru, India

Aug 2019 – May 2023

B.E. in Electronics & Telecommunication

Final-year projects in ML and blockchain.

Get in Touch

I'm open to collaborating on agent safety, evaluation methodology and regulatory NLP. Feel free to reach out if you're interested in working together — or just want to say hello.

Location

Dublin, Ireland