Every AI Agent Needs a Security Boundary

A chatbot produces an answer. An AI agent may search a drive, read email, call an external service, write to a database, send a message, create an account or trigger a transaction. The ethical question therefore changes when language generation acquires tools and authority.

An institution cannot secure an agent merely by telling it to behave. It must govern the systems, credentials, permissions, data and actions around the model. The same capability that makes an agent useful—its power to act across connected environments—also allows a mistaken, manipulated or compromised agent to spread harm faster than a human reviewer can notice.

A trustworthy institutional boundary follows a clear chain: legitimate purpose → distinct agent identity → task-bound authority → minimum data and tools → independent approval for consequential action → verified execution → complete logs → containment → remedy → revocation and recovery.

The goal is neither fear nor friction for its own sake. It is useful autonomy without invisible power.

Reuters reported at 03:04 East Africa Time on 8 October that CrowdStrike had linked an AI-assisted campaign against South Korean financial organisations to infrastructure containing an agentic penetration-testing tool and coding-model session records. At least nine banks had reportedly been targeted since late September. One bank disclosed personal-information exposure affecting about 25,000 customers. Reuters

CrowdStrike's underlying report says investigators found open directories containing Claude Code…