AI agents are moving quickly from demos to daily work. They read inboxes, raise tickets, query databases, call APIs and, increasingly, take actions on their own. Each of these agents is effectively a new digital worker, and like any worker it needs an identity, permissions and supervision. Gartner lists both oversight of agentic AI and identity and access management for AI agents among its top cybersecurity trends for 2026. For most organisations, this is the next big identity challenge.
Why AI agents are an identity problem
Traditional identity and access management (IAM) was designed for people and, later, for service accounts that behave predictably. AI agents are different:
- They act autonomously. An agent decides which tools to call and in what order, based on instructions that can change at run time.
- They can be manipulated. Prompt injection hidden in an email, document or web page can redirect an agent to do something its owner never intended.
- They multiply fast. Teams can create new agents in minutes, often outside IT’s view, each with its own API keys and tokens.
- They often borrow human access. Many agents run with the full permissions of the employee who set them up, which is far more than they need.
Five controls for agent identity
1. Give every agent its own identity
Register each agent as a distinct, non-human identity with a named human owner, a purpose and an expiry date. Avoid shared credentials and never let agents run under a person’s account. This makes every action traceable.
2. Apply least privilege and short-lived credentials
Grant only the scopes an agent needs for its task. Prefer short-lived tokens issued just in time over long-lived API keys stored in configuration files, and rotate secrets automatically.
3. Put a human in the loop for high-impact actions
Define which actions an agent may take on its own and which need approval: sending external emails, moving money, changing access rights or deleting data should usually require a human decision.
4. Use policy-based authorisation
Instead of hard-coding permissions into each agent, enforce central policies that decide in real time whether an action is allowed, based on the agent, the data, the context and the risk.
5. Log, monitor and be able to switch it off
Record every prompt, tool call and action. Feed agent activity into your SOC so unusual behaviour is detected quickly, and make sure every agent has a tested kill switch.
Start with an inventory
Most organisations do not yet know how many agents, bots and automation accounts they have. A simple first step is to inventory non-human identities across cloud platforms, SaaS applications and code repositories, then remove what is unused and reduce what is over-privileged.
How A1 Technology can help
We help organisations discover non-human identities, design access controls for AI agents and monitor them alongside the rest of their environment. Our AI assurance services and information security solutions give you the visibility and control to adopt agents safely. Book a free security assessment to get started.


