cybersecurity operations center analysts monitoring enterprise network

Zscaler (NASDAQ: ZS) is working with Anthropic and Mistral on integrations that the cybersecurity company says will give enterprise defenders more visibility into AI-agent activity, attack paths and the controls around production AI use. The effort combines Zscaler’s Security for AI offering with frontier-model capabilities at a time when organizations are connecting agents to business data, applications and tools.

The collaboration has two distinct tracks. With Anthropic, Zscaler said it is participating in the company’s Cyber Verification Program to advance vulnerability discovery and security research. Zscaler also said it has integrated Anthropic’s Claude Mythos 5.1 into Endpoint AI Security to power a capability called Exploit Paths, which correlates security evidence across a sequence of events and highlights places where a defender could interrupt an attack chain.

The company plans a limited preview of Exploit Paths. That limitation matters for security teams assessing whether model-assisted analysis can be placed into an incident process. The source does not describe general availability, supported data sources, pricing or independent detection results, so organizations will need to validate how the feature performs against their own telemetry and response procedures.

The Mistral work is aimed at European customers. Zscaler said its Security for AI service, delivered through the company’s sovereign Zero Trust Exchange in Europe, will integrate with Mistral to support enterprise AI adoption under security, data-protection and sovereignty requirements. The company did not specify which Mistral models or deployment patterns the integration supports.

For IT and security operations teams, the practical issue is not only whether an agent can produce a useful result. Agents may receive access to internal knowledge, cloud services or operational tools, creating an access-management problem that resembles identity and application security but can move at machine speed. Zscaler said its platform is intended to provide controls over which tools and applications an AI agent can reach, rather than giving an agent broad network access.

The Exploit Paths feature is positioned as a way to use model reasoning within an investigative workflow. A security operation often accumulates evidence from endpoints, identities, cloud services and network activity, then asks analysts to determine whether separate signals describe a connected intrusion. Zscaler said the capability is designed to correlate that evidence and identify potential threats and attack paths. That is a vendor-described function, not an independently established measure of detection quality or remediation speed.

The announcement also illustrates how AI security is becoming a platform-integration question. Enterprises can adopt a frontier model through a direct API, a cloud service or a specialized application, but each route raises questions about identity, permissions, sensitive-data handling and monitoring. A security layer that can apply consistent controls across those routes could reduce operational fragmentation, although the source provides no customer deployment data to demonstrate that outcome.

Legacy perimeter tools and virtual private networks are typically designed around users and network segments, not autonomous software that calls multiple tools during a task. Zscaler is positioning its zero-trust architecture as an alternative that connects users, workloads, devices and AI agents only to authorized applications and tools. Whether the integrations improve a particular organization’s defenses will depend on how it defines agent permissions, integrates security evidence and retains human review for consequential actions.

For teams moving AI agents from pilot projects into production, the immediate value is a more concrete security design to evaluate: model-assisted attack-path analysis paired with access controls and regional deployment considerations. The company’s limited-preview timeline means those teams should treat the new Anthropic capability as an evaluation target rather than an established operational control.

Leave a Reply

Your email address will not be published. Required fields are marked *

Latest News