Varun Iyer
Hi, I'm Varun. I research the tools that catch dangerous AI behavior. My work shows these monitors can pass every test and still be broken. Today's safety monitors fail silently — and no one catches it.
Previously: co-founder / CTO of three startups — most recently Soma (decentralized foundation-model training), Glass (decentralized video, made $1M+ for creators), and Spott (map-based social network). Before that, an undergrad at UChicago working on neuroscience-inspired AI and a math REU on search engine algorithms.
Recent writing
-
Linear Safety Probes Cannot Silence Features They Detect
You can build a safety monitor that perfectly spots dangerous behavior — then disable it, and the model does the dangerous thing anyway.
-
Fine-Tuning Silently Breaks AI Safety Monitors
Fine-tune a model and its safety monitor quietly stops working — even though every standard test still says it's fine.