'Near-autonomous' AI agents attack Taiwan's nuclear safety agency

Researchers observed a multi‑wave, agentic intrusion (July 1–4) in which AI agents built on Hermes and OpenClaw autonomously mapped a government portal, discovered unauthenticated APIs and exposed user databases, solved CAPTCHAs, conducted password‑spray attacks to compromise 85 accounts, and exfiltrated ~2,564 personnel records plus SSO client secrets and database credentials; the framework then scanned and attacked supply‑chain vendors, a nuclear safety agency, a government email system, and multiple energy companies while using autonomous learning cycles and self‑correction.