Wiz Launches AI Security Agent to Find Hidden Risks in Enterprise AI Apps

Wiz's new AI Security Agent analyzes application code and cloud permissions to uncover hidden AI agents and potential attack paths. The tool is in public preview, with discovery improvements reported by Wiz but not independently verified.

Nov 18, 2026
4 minute read
eSecurity Planet content and product recommendations are editorially independent. We may make money when you click on links to our partners. Learn More

An AI assistant built to answer internal questions could also have permission to execute code, modify files, or access sensitive cloud data. Those capabilities may go unnoticed when security tools only examine the infrastructure hosting the application.

Wiz launched its AI Security Agent in public preview on Nov. 16, 2026, to address that visibility gap. The tool analyzes AI applications and agents, identifies their intended purpose, and compares it with their actual capabilities and cloud permissions.

Wiz says the tool has discovered up to five times as many AI agents as traditional discovery engines across customer environments. The company hasn't disclosed its testing methodology or false-positive rates, so the figure remains a vendor-reported result rather than an independently verified performance measure.

How the Wiz AI Security Agent discovers hidden risks

According to Wiz's announcement, the AI Security Agent extends the company's cloud security capabilities into application code and AI-specific configurations.

Advertisement

Conventional discovery methods may identify agents registered with known platforms or exposed through management APIs. Custom agents embedded in application code can be harder to detect.

Wiz says its new tool analyzes three environments:

  • Self-hosted workloads: AI applications running on virtual machines, containers, serverless infrastructure, and code stored in buckets.
  • Supported PaaS platforms: Custom AI applications defined in code.
  • Code repositories: Application logic, prompts, and configurations during development.

It inventories models, system prompts, Model Context Protocol (MCP) servers, tool definitions, and other components even when no central management API exists.

The tool then examines application instructions, logic, and permissions to determine what an agent was designed to do and what actions its configuration permits.

An internal research assistant, for example, might have unnecessary access to external networks or infrastructure commands.

Wiz says the findings include explanations and source-code references so security teams can investigate why a capability was identified.

Advertisement

The approach addresses risks already documented in attacks against MCP servers and AI infrastructure, where connected tools and credentials can extend the consequences of a compromise beyond the initial application.

From application code to cloud attack paths

Wiz combines its AI application analysis with the Wiz Security Graph, which maps infrastructure, identities, permissions, exposure, and sensitive data.

In one example presented by the company, an internal AI assistant contains an unvalidated query entry point that accepts arbitrary destination URLs.

Wiz says its analysis identified a way this behavior could expose a backend service account's OAuth token through an authorization header.

The Security Graph then connected the finding to two internet-reachable endpoints, an overprivileged service account, and storage buckets containing personal and financial information.

Although the assistant had a guardrail enabled, Wiz found that the guardrail did not prevent access to the connected storage.

The example demonstrates how a weakness in application logic can combine with cloud permissions to create a broader attack path.

Advertisement

It does not establish that the illustrated weakness was exploited by an attacker or that similar configurations exist across every customer environment.

What security teams should check before using the preview

Wiz says the AI Security Agent can identify risks arising from combinations of instructions, tools, and permissions that traditional rule-based scanners may miss.

For example, an unauthenticated endpoint accepting attacker-controlled instructions could become a remote-code-execution risk if the connected agent can execute shell commands.

Such weaknesses can also appear in indirect prompt injection attacks, where untrusted content influences an agent's behavior.

However, discovering a potential attack path is not the same as confirming exploitation.

Organizations evaluating the preview should validate its findings against their actual application configurations and enforce access restrictions independently of the agent's instructions.

Advertisement

Key questions for a pilot deployment

Before expanding the tool across an organization, security teams should establish:

  • Discovery coverage: Does it identify custom agents across supported repositories, workloads, and PaaS environments?
  • Accuracy: Which findings are confirmed, and how many require manual review?
  • Access requirements: What permissions does the scanner need to inspect application code and cloud resources?
  • Risk validation: Can teams reproduce the reported permissions or exposure without affecting production systems?
  • Remediation: Who owns the affected agent, and which access changes or code fixes are required?

Security teams should also compare the tool's inventory with existing cloud and identity records to identify agents that other discovery methods missed.

That review complements broader AI agent access controls intended to restrict unnecessary permissions and prevent unauthorized actions.

Advertisement

What is available now?

The Wiz AI Security Agent is available through the Wiz Preview Hub.

Its announced capabilities include agent discovery, application-purpose analysis, and risk correlation using cloud context.

Wiz describes future runtime context integration as a way to continuously validate agent behavior during execution. That should not be treated as a confirmed capability of the current preview.

The company also refers to additional specialized code-to-cloud and AI static-testing capabilities. Separately, Wiz AI SAST was already introduced in public preview in October 2026.

Wiz hasn't published independent benchmark results for the AI Security Agent's discovery accuracy or comprehensive information about preview limitations.

For organizations adopting agentic applications, the immediate priority is establishing which agents exist, what resources they can access, and whether those privileges are necessary.

Wiz's public preview offers another way to investigate those questions. Its value will depend on the accuracy of its findings and whether security teams can use them to identify and reduce verified risks.

Advertisement

Also read: Researchers found that AI agents could bypass spending limits when approval codes were accessible through their connected email accounts.

eSecurity Planet Logo

eSecurity Planet is a leading resource for IT professionals at large enterprises who are actively researching cybersecurity vendors and latest trends. eSecurity Planet focuses on providing instruction for how to approach common security challenges, as well as informational deep-dives about advanced cybersecurity topics.

Property of TechnologyAdvice. © 2026 TechnologyAdvice. All Rights Reserved

Advertiser Disclosure: Some of the products that appear on this site are from companies from which TechnologyAdvice receives compensation. This compensation may impact how and where products appear on this site including, for example, the order in which they appear. TechnologyAdvice does not include all companies or all types of products available in the marketplace.