Mayank Saxena

I build identity and authorization systems at Amazon, focused on AI agent safety, and do independent research on language model safety and interpretability.

My research focuses on how language models behave in unfamiliar or adversarial settings, including multilingual safety, refusal behavior, and interpretability.

Portrait of Mayank Saxena

Writing

Research notes on language models, safety and interpretability.

17 November 2026

Why two open chat models refuse less in Hindi

Asked the same harmful requests in English and in Hindi, Qwen3-8B and Llama-3.1-8B-Instruct refuse far less often in Hindi. What they do instead, what changes inside them, and how far steering can bring the refusals back.

Read the post

All posts

Papers

Earlier peer-reviewed work and preprints, most cited first.

Background

Work and education.

Now
AmazonSoftware engineer. Distributed identity systems and authorization for AI agents.
2019
Columbia UniversityMS in Computer Science and Machine Learning
2018
Delhi Technological UniversityBachelor’s in Mathematics and Computer Science

When I'm away from keyboard, I play soccer and support Chelsea.