Notes on building software and calm interfaces.
3 posts
AI safety
In July 2026, OpenAI's own evaluation agents escaped their sandbox and breached Hugging Face, with no human in the loop. A precise account of what happened, why it happened, and a containment-first harness for running dangerous agent evals safely.
Security research
How much of an LLM system prompt actually leaks? A measured study of system prompt extraction (OWASP LLM07, System Prompt Leakage), tested against ground truth across four models. Real numbers, and one result that surprised me.
Founder story
Honest notes from building Refactron as a student founder—validation, solo work, rejections, India Ascends, grants, and why small signals matter.