Under 20 Prompts Found Zoom's Flaw. OpenAI Sells the Defense.

Three items landed in the same 24-to-72-hour window that describe the same thing from different angles: the cost of finding things has collapsed. Taiwan's nuclear regulatory agency was breached in a campaign the Financial Times describes as AI agents running reconnaissance and intrusion in parallel rather than assisting a human operator. Researchers say they surfaced a Zoom device-takeover flaw using fewer than twenty prompts on publicly available models. And OpenAI shipped GPT-5.6-Cyber into AWS for authorized defenders — a commercial defense sold against a threat that has, per the reporting, already been demonstrated in the field. The same dynamic shows up outside security: an unreleased research version of Claude raised a proven bound on a Riemann-related problem from 4 ...