AI Finds and Exploits Zero‑Days

OpenAI reports that a GPT‑6 'Cyber' model helped an automated system identify 1,724 Linux kernel flaws and use 948 of them to gain admin access in testing, and it uncovered dozens of previously unknown high‑severity browser vulnerabilities. Human researchers reviewed the results and filed disclosures, but the experiment shows models can autonomously produce novel exploit techniques at scale. The testing reportedly consumed about $1.2M in API credits, indicating real economic and operational feasibility. — This demonstrates a new security reality: general‑purpose AI can discover and weaponize zero‑day vulnerabilities, creating urgent questions about lab safety, disclosure norms, liability, and whether/ how such research should be governed.