Wiz Launches AI Security Agent to Find Hidden AI Agents and Map Their Attack Paths
Wiz Research has announced the Wiz AI Security Agent, now in Public Preview. It is built around a simple problem: AI agents work at the application layer, orchestrating prompts, calling tools and running business logic, while traditional cloud security looks at identities, workloads, permissions and data. Wiz argues that security teams need to see both layers and understand how they connect.
Wiz gives a hypothetical example. A BigQuery agent has tools with read and write access to several databases, and its code contains an arbitrary code execution flaw. On its own, the code finding is only half the story. By feeding it into the Wiz Security Graph, the tool checks external exposure, internet egress and effective BigQuery permissions. That shows how an outside attacker could hijack the agent's tools and reach sensitive data.
Discovery is a major focus. Custom agents defined entirely in code may have no management API or recognisable signature, so conventional inventories can miss them. Wiz says the agent inventories AI applications, models, system prompts, MCP servers and tool definitions across self-hosted environments (VMs, containers, serverless and code in buckets), supported PaaS platforms and code repositories. Across many customer environments, Wiz reports it detected up to five times as many AI agents as traditional discovery engines.
The tool also compares each agent's intended purpose with what it can actually do. An agent built to answer questions from internal business data might also be able to execute code, modify local files or contact external systems. It can also flag internal agents that are publicly exposed or hold data access beyond their scope. Wiz says further risk areas, including code-to-cloud analysis and AI SAST, will follow.
Why it matters
Many organisations assume that cloud asset inventories and management APIs show every AI workload, but agents embedded in application code can sit outside both, and Wiz's claim of finding up to five times more agents suggests the gap may be large. The source also shows that a code flaw such as prompt injection only becomes a real risk when paired with exposure, egress and permissions, so scoring either layer alone misjudges priority. Leaders must now treat an agent's purpose as a baseline for least privilege, not just its technical configuration.
What to do
- Inventory AI agents, models, system prompts, MCP servers and tool definitions across code repositories, PaaS and self-hosted environments.
- Review each agent's tools and cloud permissions against its intended purpose, and remove unnecessary code execution, file modification or external communication abilities.
- Check whether agents meant for internal use are exposed to the internet or hold access to sensitive data beyond their scope.
- Review agent code for prompt injection flaws that could let an attacker run arbitrary commands.