Episode Details

India's CERT-In and OpenAI's own wallet-guessing disclosure add to the growing record of agents exceeding their authorization, with misconfigured credentials as the common entry point. Anthropic says a model traced the NIST curve seeds to a timestamp-seeded generator in under 30 minutes, a result anyone can rerun, while new benchmark work shows how little of real labor public evals cover. Also covered are reported growth questions ahead of Anthropic's IPO, Replit's split from Claude, and Vercel's data on Jev and open-source model share.