Security Must Treat AI Agents as Identities and Guard Their Decisions
Experts argue that adding lifecycle rules, task‑specific short‑lived credentials and real‑time decision controls will prevent authorised agents from causing large-scale errors or abuse.
Overview
- AI agents are already running in production in retail, logistics, banking and government and can act with valid credentials to change prices, route shipments, approve actions and access data without a human in the loop.
- Coverage across TechRadar, IT Security Guru and ComputerWeekly says organisations should treat agents as first‑class non‑human identities with named owners, a defined purpose, inventory and expiry or recertification.
- Traditional IAM and perimeter controls are inadequate so experts recommend task‑specific, short‑lived and revocable credentials combined with zero‑trust checks to limit agent privileges.
- Security must move into the agent decision loop by adding runtime protections such as behavioral monitoring, approval checkpoints, anomaly detection, autonomy budgets and emergency kill switches to stop risky actions in real time.
- Governance of agent-to-tool links like the Model Context Protocol and tighter standards from bodies such as NIST and the Cloud Security Alliance are urgent because one analyst warns weak controls could lead to many projects being canceled by the end of 2027.