Bringing Anthropic's Claude Mythos 5.1 to Zscaler Endpoint AI Security

AI adoption inside the enterprise is increasingly bottom-up: employees are trying the latest AI assistants, coding agents, browsers, and development tools because these products help them get work done. Many of these tools arrive before central security or IT teams have evaluated them.

The result is a sprawl of shadow AI: a fast-growing layer of applications, agents, extensions, models, packages, and configuration files on employee devices. This activity changes quickly and often falls outside the software rollout processes organizations use for traditional applications.

Today, Zscaler is announcing the limited preview availability of Anthropic's Claude Mythos 5.1 capabilities in Zscaler Endpoint AI Security. Mythos 5.1 powers Exploit Paths, helping security teams connect related endpoint AI activity, identify potential attack paths, and determine where to intervene. 

With this addition, Zscaler better helps security teams turn disconnected endpoint events from different AI apps, across all users, into an attack chain analysis view of potential threat vectors. By correlating multi-stage AI behaviors into visible attack paths, security teams can leverage intelligent prioritization to proactively block risky agent activities, secure sensitive data, and enforce real-time policy controls directly at the endpoint.

Why MDM and EDR leave a visibility gap

MDM and EDR remain essential parts of endpoint security, but they were built to answer different questions. MDM helps organizations manage applications, devices, and configurations. EDR monitors processes, files, network activity, and known attack behavior.

AI agents introduce a different kind of endpoint activity. They can act on a user’s behalf using the user’s identity and permissions, call tools, access files, run commands, and connect to services as if the user performed those actions directly. As a result, audit logs may attribute agent activity to the user’s account, making it difficult to separate what the user did from what an agent did on the user’s behalf.

At the process or file level, this activity can look normal. A signed application may modify a file, run a command, or send data over the network without matching a malware pattern. The missing context often lives inside the AI application layer: which agent acted, what task it was performing, which tool it called, what data it used, and which extensions, models, packages, or connected services were involved.

Endpoint AI Security was built to close this gap. It secures the AI activity layer on the endpoint — where AI apps, browsers, assistants, coding agents, and other agentic tools are creating new risks that traditional endpoint tools were not designed to interpret or control.

How Zscaler Endpoint AI Security closes the gap

Zscaler Endpoint AI Security is designed to secure this AI application layer across users and managed devices. It combines all three pillars critical to securing employee devices using AI: discovery, runtime visibility, and policy enforcement.

Discover AI assets

Security teams first need to answer a basic question: What AI is running across the organization?

Endpoint AI Security identifies AI assistants, agents, browsers, development tools, models, packages, extensions, and connected services used across managed devices. This provides visibility into AI assets that may not appear in traditional application inventories.

See what agents are doing

Inventory shows what is present. Runtime visibility shows what is happening.

Endpoint AI Security captures the events that matter during an agent run, including prompts, tool calls, file access, commands, system interactions, and network activity. It connects those events to the user, agent, and device involved.

Enforce enterprise AI policy

Visibility becomes useful when it can change an outcome. Endpoint AI Security provides a policy layer for controlling AI activity on the device.

Policies can restrict unauthorized agents and tools, stop dangerous commands or instructions, and protect sensitive files and data. This gives security teams a way to support AI adoption while keeping enterprise rules in effect at the endpoint.

The challenge of connecting activity at enterprise scale

Discovery and runtime monitoring produce detailed evidence. That evidence is distributed across users, devices, agents, tools, and individual runs.

The challenge is understanding how these activities connect. Broad agent permissions, use of a third-party tool, or access to a sensitive file may each appear manageable in isolation. When these conditions occur in the same sequence, they can combine into a more serious risk.

Reviewing each event separately makes that relationship easy to miss. Security teams need a way to cross-correlate the activity, recognize possible attack progression, and determine which response will reduce the most risk.

What Mythos 5.1 adds

Mythos 5.1 powers a critical capability within Endpoint AI Security, Exploit Paths, which cross-correlates evidence to identify potential threats, explain how the steps connect, and show defenders where to interrupt the sequence.

Each path shows the evidence behind the analysis, how confident the assessment is, and where a step is inferred rather than confirmed. It also identifies how a defender could recognize the activity and where to break the chain of attack.

This adds context that individual alerts cannot provide. Security teams can categorize risk based on the complete path, understand why a group of findings matters, and prioritize the control that can interrupt the sequence.

Endpoint AI Security and Mythos 5.1 have distinct roles. Endpoint AI Security discovers AI assets, records agent activity, and provides the policy enforcement point. Mythos 5.1 analyzes that evidence, connects related conditions, and identifies the attack paths that deserve attention.

What this looks like in practice

Consider an AI agent with broader permissions than its assigned role requires. During a run, it calls a third-party tool and accesses files containing sensitive information. Endpoint AI Security records the agent, the user and device involved, the tool call, and the file activity. 

Viewed in isolation, these events may not reveal the full risk. Mythos 5.1 correlates the findings to determine whether they form a credible sequence. Exploit Paths shows how the agent’s permissions, use of a third-party tool, and access to sensitive files could combine into an attack path. The report presents the supporting evidence, identifies any inferred steps, and recommends where defenders should intervene.

The security team can then use Endpoint AI Security policy controls to restrict the agent or tool, stop risky activity, or block access to sensitive data. Mythos 5.1 provides the analysis and prioritization, while Endpoint AI Security provides the visibility and enforcement.

Turning evidence into action

Exploit Paths is designed for defensive assessments of environments customers are authorized to protect. It presents prioritized findings without exposing raw model responses or generating working exploit code or payloads. Security teams review the evidence and decide how to respond. By bringing Mythos 5.1 into Endpoint AI Security, Zscaler helps security teams turn disconnected endpoint events into a focused view of potential attack paths.

“AI agents are rapidly becoming part of everyday work. But they also introduce a new layer of endpoint activity that security teams need to understand and control. By embedding Mythos 5.1 in Zscaler Endpoint AI Security, we are helping customers connect agent behavior, tool use, and data access into clear exploit paths so defenders can prioritize the actions that reduce the most risk.”
— Dhawal Sharma, EVP AI Security, Zscaler

“Bringing Claude Mythos 5.1 into Zscaler's Endpoint AI Security helps enterprise security teams reason over complex AI activity in a defensive, evidence-based way. This makes it easier for organizations to adopt AI tools confidently while maintaining the visibility and safeguards they need.”

– Ash Alhashim, GTM Leadership Cybersecurity, Anthropic

With this addition, security analysts can see how AI risk develops across agents, tools, and data on the endpoint. As those signals come together, Exploit Paths shows where risk is forming and where intervention will matter most. With proven Zscaler policy controls, teams can interrupt risky activity before it causes business impact. The result is a safer path to enterprise AI adoption - with visibility, control, and protection on every managed employee device.

​​To learn more about Zscaler Endpoint AI Security or request a demo, visit https://www.zscaler.com/request-a-demo..