Security teams must model agents as persistent operators that can revise tactics, not as faster scripts.
Security researchers reported a campaign in which multiple AI agents were used to map networks, research vulnerabilities, attempt intrusions and change tactics when an attack path failed. The operation reportedly ran for several days and compromised government accounts in Taiwan.
The distinction from ordinary automation is persistence. A script follows a prepared sequence; an agent can choose the next action from new evidence and coordinate several lines of effort at once. That compresses work that previously required a team of operators and makes low-cost campaigns harder to distinguish from sophisticated ones.
The report is based on an external investigation and attribution remains contested, so its claims should be read with appropriate caution. The defensive lesson does not depend on attribution: agent identities need least-privilege access, short-lived credentials, full action logs and containment designed for adaptive behavior.
This briefing summarizes reported facts and adds independent context. It does not reproduce the source article's wording or structure.