Zero Networks launches AI agent least agency enforcement

Zero Networks has launched Least Agency Enforcement for AI agents, applying an OWASP security principle to enterprise AI deployments.

The product is aimed at companies giving internal AI agents access to systems, tools and administrative functions while trying to limit the security risks that can follow. It is designed to restrict what an AI agent can access, which resources it can use and when it must involve a human before taking sensitive actions.

The launch reflects a broader shift in cybersecurity as companies move from experimenting with generative AI to embedding software agents in business operations. Those agents are increasingly used to query internal data, trigger workflows, connect to cloud services and interact with infrastructure, creating a new access management problem for security teams.

According to Zero Networks research, nearly 80% of enterprises have already deployed internal AI agents, while about two-thirds do not yet have governance policies covering them. The company argues that gap leaves organisations with a growing attack surface if agents are manipulated, over-permissioned or used as a path into other systems.

Least agency

Zero Networks has tied the launch to the OWASP principle of Least Agency, which calls for limiting an agent's autonomy, access to tools and decision-making authority. The aim is to reduce the impact of prompt injection, privilege abuse and AI agent compromise by ensuring an agent cannot move far beyond its intended role.

In practice, the system uses identity-based microsegmentation, automated policy setting and just-in-time multi-factor authentication for privileged access. AI agents are therefore meant to communicate only with authorised systems, use approved resources and seek human approval before carrying out sensitive administrative work.

The controls can be applied across cloud, on-premises, Kubernetes, IoT, OT and hybrid environments. Zero Networks is positioning the product as an enforcement layer rather than a documentation framework, arguing that written governance rules alone are not enough when agents operate inside production systems.

Security vendors have increasingly focused on the possibility that AI agents could become a new route for lateral movement within corporate networks. If an agent has broad permissions and is deceived by a malicious instruction, it could touch multiple systems before a human notices. Containment has therefore become a central theme in product design across the sector.

Least Agency Enforcement is built on Zero Networks' existing AI security offering, which includes AI Agent Control, AI Segmentation, AI SaaS Control and measures for enterprise large language model deployments. The company's broader argument is that organisations need technical restrictions around machine identities and agent behaviour, not just policy documents and approval processes.

The launch also shows how established ideas in identity and network security are being adapted for AI systems. Least privilege has long been a standard principle for human users and service accounts, with the goal of limiting access to what is necessary for a task. Vendors are now trying to apply the same logic to software agents that can make decisions and chain together multiple actions with little direct supervision.

"Least privilege works because it is simple: give people access to what they need, nothing more. We're doing the same thing for AI agents, except now it must be automatic, because nobody has time to babysit a thousand agents by hand," said Benny Lakunishok, Chief Executive Officer and Co-founder of Zero Networks.

He said the same principle should apply when an agent is manipulated or misused.

"If an agent gets fooled or misused, it should hit a wall almost immediately, not wander around the network looking for something valuable. That's the bet we're making: less freedom for the agent now, which beats explaining a breach later," Lakunishok said.

Least Agency Enforcement is available now.