AI agents are moving beyond chat interfaces and into enterprise workflows, where they can access applications, handle internal data and take action on behalf of employees.
That creates a practical question for IT and security teams: when an agent behaves unexpectedly, when should it be stopped?
Andy Ulrich, Chief Information Security Officer at Vonage, told UC Today that organisations do not need an entirely new security playbook for AI. But they do need to apply established security disciplines rigorously as agents become more capable and autonomous.
“AI is controllable, just like any other threat is controllable,” Ulrich said. “You have to start with the basics.”
Those basics include identity and access management, patching, incident response and logging. They have long been central to enterprise security, but AI agents can use systems more quickly and more thoroughly than an individual employee.
Ulrich said organisations should think of an agent as an exceptionally capable employee that may not naturally recognise the limits of internal policies.
“If you consider AI to be an employee who’s super smart, super efficient, super fast, and doesn’t have strong regard for your policies, that’s sort of how we have to think about these AI agents,” he said.
Treating AI Agents as Managed Identities
AI agents can be managed in different ways as businesses test where and how they can safely operate.
Ulrich said organisations may initially want to treat AI as a privileged user, even where the employee deploying the agent would not normally have privileged access. That could give IT and security teams greater oversight while they develop a clearer understanding of agent behaviour.
However, he does not believe that approach is sustainable over the long term.
Instead, Ulrich said organisations should work towards giving an agent the same permissions as the individual responsible for it. Those permissions should then be reviewed through the same processes used for human identities.
“We should be driving towards managing what AI can access in the same way that we manage it for people,” Ulrich said.
That means maintaining visibility over who is responsible for an agent, which applications it can access and why that access has been granted. It also requires regular access reviews, much like those already carried out for employees.
Logging Must Support Security Investigations
The risk grows when an AI agent is assigned a task but does not have the permissions required to complete it. Ulrich said agents may try different routes to achieve their objectives, including seeking ways to gain access they were not originally granted.
