Kunal Singh // technical AI governance researcher

Investigating how well AI safety evaluations hold up under scrutiny — where model assessments can be trusted and where they fall short. Ongoing research under MATS 10, Technical AI Governance stream.

RED30 — 30 universal AI red-line indicators across 4 risk categories, built by analyzing 16 frontier models and mapping EU/global regulatory frameworks onto operational safety criteria. Submitted to the Apart Research Technical Governance Hackathon.

Cofounded a coordination hub connecting India's AI Safety researchers, organizations, and policy practitioners — building community infrastructure and adapting global AI safety frameworks to India's policy context.

Designed adversarial audit scenarios (vulnerability classification vs. deliberate misclassification) to test three frontier models for strategic deception under oversight — modeling third-party AI assurance methodology.

Structured red-teaming framework across 100 adversarial prompts and 6 attack categories, with an automated 0–3 compliance scoring pipeline — revealing a 69pp security gap and 51% capability exploitation rate.

Argues US export controls fail because they exempt publicly released models — anyone can fine-tune open-weight models for harmful use at a fraction of training cost. Proposes focusing controls on compute, transparency, and developer liability instead.

A comprehensive review of fine-tuning strategies for large language models, published at the 20th International Conference on Computing and Machine Learning (CML), March 2024.