The Pulse
Wiz puts its AI Security Agent in public preview
The Wiz AI Security Agent is available in public preview through the Wiz Preview Hub. It is designed to find AI applications and assess their code, capabilities and cloud permissions for potential attack paths.

Security teams can now preview a Wiz agent designed to find AI applications and assess whether their capabilities and cloud access create security risks. Wiz says it connects analysis of application code and instructions with cloud permissions, exposure and sensitive data in its Security Graph.
Source wiz.io
The Wiz AI Security Agent is available in public preview through the Wiz Preview Hub. The company says that across many customer environments it has detected up to five times as many AI agents as traditional discovery engines.
Finding agents in code and cloud
The preview builds on Wiz's existing AI discovery and cloud security capabilities. It covers AI application logic, prompts and configurations in self-hosted environments, supported platform-as-a-service platforms and code repositories.
In self-hosted environments, that includes agents and applications running on virtual machines, containers and serverless infrastructure, as well as code stored in buckets. The supported platform-as-a-service category covers custom agents and AI applications defined in code, while repository analysis covers application code, prompts and configurations during development.
Wiz says the agent inventories custom AI applications, models, system prompts, MCP servers and tool definitions, including where there is no central management API. The resulting application information is connected to the surrounding environment in the Wiz Security Graph.
Comparing purpose with access
The agent also analyzes implementation, logic, instructions and permissions to reconstruct an application's architecture, according to Wiz. It compares an agent's intended purpose with its capabilities and infrastructure access, looking for discrepancies such as an internal agent exposed externally or access to sensitive data beyond its intended scope.
One example in the company's post is an assistant intended to answer questions using internal business data that also has the ability to execute code, modify local files or communicate with external systems. The analysis includes reasoning and supporting evidence, including source-code references, so security teams can examine how capabilities were identified.
Wiz plans to add specialized risk domains including code-to-cloud analysis and AI SAST. It also describes future runtime context integration intended to validate agent behavior continuously during execution.
Tracing potential attack paths
In a hypothetical BigQuery example, Wiz describes the security agent finding tools with read and write access to multiple databases alongside an arbitrary code execution vulnerability. The company says cloud context can then be used to check external exposure, internet egress and effective permissions when assessing the potential path to sensitive data.
In another example, an internal cloud assistant has a query entry point that accepts an arbitrary destination URL and leaks backend service-account OAuth tokens in an Authorization header. Wiz's example traces two internet-reachable endpoints upstream of the agent to an overprivileged service account and storage buckets containing names, emails and financial data.
The product provides findings with explanations, potential impact and supporting evidence, according to the company. Combined with cloud context, those findings are intended to help teams understand where application-level weaknesses intersect with exposure, identities, permissions and sensitive data.
Source
Companies in this story
Related stories
- Vercel opens paid Security Review to unlinked GitHub repositories
- Cisco Talos warns AI pentesting tools can speed intrusions
- LangChain puts agent records at the center of LangSmith
- ChatGPT Wallet lets US Pro subscribers approve agent purchases
- MatrAIx opens proposals for long-horizon personal agent benchmark