Article written by Jonathan Perz, Manager of Information Security at Abacus Technologies
The security conversation changes when artificial intelligence stops answering questions and starts taking actions.
An AI assistant might summarize an email, analyze a spreadsheet, or draft a response.
An AI agent may read the email, access the spreadsheet, update a business system, create a ticket, send the response, or initiate another workflow.
That difference matters.
The core cybersecurity issue with AI agents is not simply intelligence.
It is authority.

What Is the Agent Allowed to Do?
Every action an AI agent performs depends on some form of access.
The agent may inherit a user’s permissions. It may use an application identity. It may connect through an API, OAuth authorization, service account, connector, or another integration.
Those permissions determine the potential impact if the agent makes a bad decision, receives malicious input, is misconfigured, or is manipulated.
Consider the difference:
An AI assistant says:
“This invoice appears legitimate.”
An AI agent says:
“I reviewed the invoice, approved it, updated the accounting system, and notified the vendor.”
The second scenario introduces a much larger security question.
Who authorized the agent to approve the invoice?
What prevented it from exceeding that authority?
Could malicious input influence the decision?
Could the agent access information unrelated to the task?
Would the organization know exactly what the agent changed?
AI Agents Need Traditional Security Controls
AI may be new, but many of the security principles are not.
Organizations deploying AI agents should still apply:
- Least privilege.
- Separation of duties.
- Strong identity controls.
- Limited and reviewed application permissions.
- Logging and audit trails.
- Approval requirements for sensitive actions.
- Defined boundaries on what an agent may access or change.
- Regular review of connected applications and integrations.
The difficult part is that AI agents can operate across multiple systems and make decisions dynamically.
That makes access that once seemed harmless considerably more powerful.
Organizations should therefore evaluate an AI agent based not only on what it is designed to do, but also on everything it technically could do if something went wrong.
As AI moves deeper into business processes, this becomes one of the most important questions in AI cybersecurity:
If an AI agent took an unauthorized action tomorrow, could you determine what authority it had, where that authority came from, what it accessed, and exactly what it changed?
If not, the problem is bigger than AI.
It is an access control problem wearing an AI label.