Netskope launches control to block risky AI agent actions

Netskope has launched Skylight Agent Action Control, part of its newly renamed Skylight AI Security suite.

The product is aimed at security teams that want to block risky AI agent actions before they happen, rather than investigate after an incident.

Netskope also used the launch to rebrand its broader AI security portfolio. Previously known as Netskope One AI Security, it now carries the Skylight name and includes AI Command Centre, AI Guardrails, AI Gateway, Agentic Broker and AI Red Teaming.

Agent Action Control classifies each attempted AI agent action before execution and applies policy controls based on that classification. The tool sorts actions into nine intent-based categories: access control changes, configuration changes, credential and secret manipulation, data destruction, infrastructure provisioning, potential data exfiltration, potential external communication, remote code execution and source code changes.

Security teams can then set policies to block, allow or alert based on risk levels marked low, medium, high or critical. Policies can be applied to individual agents, allowing a coding assistant to be governed differently from a chat application.

All actions are logged, enabling teams to filter alerts by areas such as cost exposure, source code changes, infrastructure updates or external communications. The feature works on network traffic already inspected by the Netskope platform, so customers do not need a separate console or an additional agent.

In one example, a coding agent attempts to delete a production repository after receiving a vague prompt. The action would be classified as data destruction at a critical risk level and blocked before it reached the repository.

Market pressure

The launch comes as companies face growing pressure to govern autonomous AI tools more closely. Netskope cited research showing that 91% of organisations cannot stop a risky agent action before it executes, while 54% reported a confirmed or suspected AI agent security incident over the past year.

The issue centres on what Netskope calls "authority drift", where agents operate with permissions or instructions that become too broad or poorly defined. That can leave organisations exposed when agents move from assisting users to making changes in systems, code bases or infrastructure.

John Martin, Chief Product Officer at Netskope, described the product as a way to put policy checks in front of each action taken by an AI agent.

"AI agents tend to act first and explain later, and most security teams only learn what happened after it is done," Martin said.

"Agent Action Control puts a decision in front of every action an agent takes, so a team can say yes to agentic AI without saying yes to the one action that could cost them a production system," he added.

Broader shift

The move reflects a broader change in enterprise security thinking as companies experiment with AI agents that can act on behalf of staff. While many existing controls focus on monitoring prompts, outputs or model use, security teams are increasingly concerned about whether agents can alter systems, move data or execute code without enough oversight.

That has pushed suppliers to frame AI security less around model safety alone and more around operational controls tied to business systems and workflows. Netskope's approach places the control point at the attempted action itself, rather than solely at the user prompt or application layer.

Industry analysts have also pointed to the limits of statistical or behaviour-based approaches when agents are allowed to interact directly with sensitive environments.

"As enterprises accelerate adoption of AI agents, we're finding that probabilistic controls are sometimes insufficient to protect the enterprise, but even occasional failure is unacceptable," said Dr. Grace Trinidad, Research Director for AI Security and Trust at IDC.

"These hardened, policy-based, deterministic controls are the backstop that prevents AI agents from causing an enterprise incident," Trinidad said.

Netskope said Agent Action Control is due to become available by the end of the current quarter.