Divij Chawla

cs + physics @ uw · spar research fellow · ai security & evals

bio

written by me

I study computer science and physics at the University of Washington. I work on AI security and evaluation: right now as a SPAR research fellow, building a benchmark for whether AI agents resist being shut down. Before that, red teaming and safety evals at Walled AI (first author, EMNLP 2025), and backdoor research at the University of Bristol.

formal casual
humble hype
literal poetic
steering vector · layer 14 of 32 · 960 dims ‖v‖ = 0

model asleep. drag a dial, then wake it.

This is SmolLM2-360M running on your own machine (WebGPU, or WebAssembly as a fallback). The dials don't touch the prompt. Each one adds a direction to the model's residual stream, found by contrasting its activations on casual vs. formal, hype vs. humble, and poetic vs. literal text. Nothing leaves your browser. The text it writes is unedited and will get facts wrong; the real ones are below.

work

  1. 2026 –

    SPAR Research Fellow Supervised Program for AI Research

    Building Shutdown-Bench, a benchmark for LLM and agent shutdownability in realistic tool-use tasks: scenario suite, instruction hierarchy, and failure-mode taxonomy. Red-team agent harnesses with provider SDKs, plus trace scoring for delay, deflection, evasion, and goal preservation.

  2. 2025

    AI Security Researcher Walled AI · Singapore

    First author, EMNLP 2025 Industry Track. Built FinRiskEval: 1,720 profiles, 8 models, 13k+ outputs measuring correctness and intent alignment in finance. Co-authored IMDA-commissioned LLM deployment playbooks on red teaming, safety evals, and guardrails.

  3. 2024 – 25

    AI Research Intern University of Bristol

    Eval harnesses for latent alignment failures via representation backdoor attacks. Fine-tuned attacker and defender models and measured sleeper-agent persistence under white-box and black-box settings.

  4. 2023 – 25

    Security & DSP Intern Emsec · Bengaluru

    Ran a honeypot fleet simulating 7k+ vulnerable apps and turned attacker telemetry into detection signals. Optimized C++ DSP modules for high-throughput SDR pipelines.

school

University of Washington, Seattle
B.S. Computer Science, double major in Physics · 2028

elsewhere

Interactive Intelligence (I2), UW
Husky Satellite Lab