Wiz Survey: 55% of Firms Run AI Agents, Only 17% Track All AI Assets

Agents are the most common AI deployment

Wiz surveyed 100 senior IT and security decision-makers, including CxOs, VPs and directors. According to the company, AI adoption is close to universal: only 5% of organizations report no active AI usage, while 55% actively deploy autonomous agentic systems. That makes agents the single most common AI deployment architecture in the survey.

Scale is also significant. Nearly a third of enterprises, 31%, run more than 25 distinct AI systems in production. Wiz says this adoption goes well beyond prompt-based tools and covers managed cloud services, proprietary in-house models, and orchestration layers such as Model Context Protocol (MCP) and RAG.

Visibility gaps and shadow AI

Security tracking has not kept up with deployment. Only 17% of organizations have a comprehensive, up-to-date AI asset inventory. The other 83% have partial visibility, fragmented team-level records, or no centralized way to catalog their running models, agents and pipelines.

The result is shadow AI. Wiz reports that 52% of organizations found unauthorized AI applications running inside their perimeters over the past year. Decentralized governance plays a part: 35% formally allow teams to deploy AI tools without centralized approval, and 9% of security leaders say they have no mechanism at all to detect unauthorized AI workloads.

What worries security leaders

When asked about agentic security concerns, respondents focused on business impact:

• Unauthorized access to sensitive data: 70%
• Unmonitored autonomous decision-making: 54%
• Overprivileged identities: 53%

Risks tied to the execution pathways themselves ranked lower. These include unmonitored agent actions (45%), lateral movement between systems (24%) and unvalidated tool APIs (23%). Wiz argues these are the routes through which overprivileged machine identities reach corporate data, and says protecting that data needs context across data posture, agent identities and runtime execution.

Asked which single AI security problem they would most like to remove, 26% named governance and cross-functional ownership complexity, narrowly ahead of agent autonomy and service account token access. Spending is fragmented: 43% of teams stretch legacy cloud tools to cover AI, 38% fund dedicated AI security budgets, and 37% rely on homegrown scripts. Wiz says this creates isolated telemetry silos that hide attack paths linking code, credentials and cloud runtime.

Wiz's recommended readiness model

Wiz says threat readiness depends on speed of action and breadth of visibility across the full lifecycle. Its framework sets out five priorities:

• Treat AI stacks as core production infrastructure, with standard configuration, patching, logging and admission controls.
• Move from flat inventories to graph context, mapping AI assets to owners and prioritizing exposure paths, for example with Wiz AI Security.
• Enforce behavioral least privilege and human oversight, restricting agent credential scopes and requiring human verification for high-impact production actions.
• Continuously test live perimeters with external AI penetration testing, such as Wiz Red Agent.
• Build a closed-loop defense using deep AI code analysis, agentic remediation such as Wiz Green Agent, and real-time detection with Wiz Defend, Wiz Sensor and Wiz Blue Agent.

What it means for security teams

The findings come from a vendor that sells products in this area, and the sample is 100 respondents, so the numbers are best read as a snapshot rather than a market-wide measure. Still, they suggest that many organizations run agents and dozens of AI systems without a full picture of what exists or who owns it. Wiz concludes that defending against machine-speed threats requires retiring manual triage loops, and that security teams need a unified code-to-cloud view to trace asset ownership and close reachable attack paths. The full survey data is available in the complete report.