AI-assisted intrusion and expanding AI-agent attack surfaces

ARCYBER has tested agents in a sandbox and allowed limited, human-reviewed network-data-platform work; a governmentwide AI-cyberattack tabletop is planned for early November. CrowdStrike reports agentic intrusions affecting at least 15 Argentine healthcare organizations and targets in other countries, but these vendor-reported claims do not establish broad autonomous capability. New commentary sharpens the distinction between securing AI systems and governing agent permissions, approvals, and traceability; guidance continues to stress ownership, stop authority, and monitoring AI-token usage.