Most enterprises securing their AI rollout are still asking the wrong question. They’re asking whether their models are accurate. They should be asking who — or what — has the authority to act on the answers.
That distinction matters more with agentic AI than it did with any AI wave before it. An AI agent doesn’t just generate a response for a human to review. It scans inventory, talks to customers, coordinates with logistics providers, and executes decisions on its own. KPMG has already tracked agentic AI deployment tripling across global organizations despite the known risks. The risks aren’t hypothetical anymore. They’re operational.
The attacks that come with this shift look different too. Prompt injection, data poisoning, model exploitation — these don’t target a software vulnerability the way a traditional exploit would. They target the reasoning of the agent itself, feeding it inputs designed to manipulate what it decides to do next. A firewall doesn’t catch that. Neither does a patch cycle.
The rise of agentic AI: Enterprise security playbook, built with its in-house ZLabs research team alongside its observability and ITSM teams, lays out a different way to think about this. Not one control, but three working together: data governance to keep the agent’s inputs clean, identity and access management to define exactly what it’s allowed to touch, and observability to catch it the moment it starts acting outside those lines.
This article was originally published on ManageEngine Academy.