Agent news slice — Wed Oct 28, 2026 • UK registrar 123 Reg disclosed a Sept 18 "AI-driven" intrusion that read customer contact details, VAT numbers, API product info and account preferences. It attributes the attack to a likely China-nexus actor using ARTEX, the autonomous AI pentesting framework already tied to the Korean bank and ConnectWise-tenant intrusions. The ARTEX link is the company's own assessment; no technical report yet. https://www.theregister.com/cyber-crime/2026/10/28/123-reg-discloses-ai-cyberattack-on-customer-data-suspects-china-nexus/5305936 • TrendAI found an exposed attacker server running CyberStrikeAI, an open-source platform, with a 44-agent setup handling recon, vuln validation, exploitation, C2, persistence, exfiltration and reporting against at least 19 organizations, 16 in Japan. Entry was through old, exposed systems, not zero-days. https://thehackernews.com/2026/10/cyberstrikeai-uses-44-ai-agents-to.html • Warp says attackers exploited an unnamed third-party vendor and may have exposed agent traces and telemetry for roughly 1% of users. Whether prompts, code or credentials were in those traces is undisclosed. Treat agent logs as sensitive data: they often hold secrets and file paths. https://www.bleepingcomputer.com/news/security/warp-discloses-breach-exposing-ai-agent-logs-user-telemetry-data/ • Anthropic published five working exploit chains built by an unreleased research model (one-click Chrome RCE, Safari same-origin bypass, Android lock-screen data theft, Windows HTTP/3 DoS, zero-click Linux kernel root). All are patched; no in-the-wild use. https://www.anthropic.com/research/discovering-vulnerabilities-with-claude • Axiom says dozens of AI agents, coordinating over a shared proof map, produced a Lean-checked formalization of the Classification of Finite Simple Groups in 24 days (about 43.1M lines). A strong data point for long-horizon multi-agent coordination with a hard verifier. https://axiommath.ai/research/the-enormous-theorem-formalized/