- Wiz polled 100 senior IT and security leaders. It says only 17% keep a comprehensive, current inventory of their AI, and 52% found unauthorized AI applications in the past year.
- About a quarter, 26%, named ownership complexity as the single headache they would remove. The margin over the runner-up was narrow, and most respondents chose something else.
- Leaders should ask who owns each AI system, what credentials its agents hold, and which actions need a human to approve them. This is a small vendor survey.
A list comes before a control
A company cannot protect a system it does not know it has. For AI, a new vendor survey suggests that is a live problem.
Security vendor Wiz polled 100 senior leaders in IT and security. They ranged from executives to directors. Only 5% said their organization has no active AI use. Another 55% run autonomous agents. These are AI programs that carry out multi-step tasks without a person directing each step.
Wiz calls agents the most common AI setup among those polled. It also says 31% run more than 25 separate AI systems in production.
Then comes the gap. Wiz says just 17% keep a comprehensive, current record of their AI assets. The other 83% have partial visibility, scattered team-level records or no central catalogue of their models, agents and pipelines.
Shadow AI is something companies found
Wiz says 52% of the organizations found AI applications nobody had approved, running on their own networks, in the past year. That is a finding, not a forecast.
Wiz links this sprawl to decentralised governance. It says 35% of organizations formally let teams deploy AI tools without central approval. It also says 9% of security leaders have no way to detect unapproved AI workloads.
Picture a company where every department can buy equipment and no central asset register exists. Finance could not say what the firm owns. That is only an illustration.
The survey, as reported, does not show why inventories are missing. It does not show whether anyone was ever asked to keep one. The 17% figure also covers all respondents, not only those with open approval policies.
How an agent becomes a route between systems
A chatbot answers a question. An agent takes actions on company data and systems. To do that, it has to log in somewhere.
Wiz points to overprivileged machine identities. These are non-human accounts with more access than their job needs. It also names service account token access as a concern. A service account is a login that a program uses. A token is the key it shows at the door.
Here is an illustration, not a case from the survey. An agent gets a token for a shared file store so it can read reports. The token was set up broadly, so it also reaches a customer database. Anyone who steals that token gets the same reach. That is lateral movement: hopping from one system to the next on access that was never needed.
Tool interfaces add a second path. A tool interface is the link an agent uses to call another program, such as a search or a database. Wiz lists Model Context Protocol (MCP) among the layers in production use. It names unvalidated tool APIs as a risk, which means calls that nobody checks before they run.
Leaders told Wiz their top worries are outcomes. Unauthorized access to sensitive data drew 70%, unmonitored autonomous decision-making 54% and overprivileged identities 53%.
The routes to those outcomes drew less attention. Unmonitored agent actions came in at 45%, lateral movement at 24% and unvalidated tool APIs at 23%. Wiz argues that these are the pathways through which overprivileged machine identities reach company data.
Ownership was the top single headache
Leaders were asked which one AI security problem they wish would disappear. About a quarter, 26%, named governance and cross-functional ownership complexity. Wiz says that narrowly beat agent autonomy and service account token access.
Wiz gives no figure for that runner-up group. So most respondents chose something else. Ownership is a leading complaint here, not a proven root cause.
In many firms, security, engineering, data and legal teams may each hold part of an AI system. That can make an owner hard to name. The survey measures inventory completeness and ownership difficulty separately. It does not report that unlisted systems have no owner.
Spending looks scattered too. Wiz says 43% stretch older cloud tools to cover AI. It says 38% fund dedicated AI security budgets, and 37% rely on homegrown scripts. Those shares add to more than 100%, so the groups overlap.
What the survey cannot tell you
Treat the numbers with care. This is a survey of 100 people, run by a company that sells security products. Respondents describe their own organizations.
The blog post shares only key takeaways. The full data sits behind a request for the report.
Wiz's five recommendations also point toward its own products. Some advice stands without a vendor: treat AI systems as production infrastructure, restrict agent credentials, and require human sign-off for high-impact actions.
Questions to put to your team this quarter
First, ask who owns each AI system running in the company, by name. If the answer takes days to find, that is a finding.
Second, ask what credentials each agent holds and whether they match its task. Wiz recommends narrow scopes to limit lateral movement.
Third, ask how the company would spot an AI tool nobody approved. Nine percent of the security leaders Wiz surveyed say they cannot.
Fourth, ask which agent actions need a person to approve them, and who that person is.
An inventory is not a security control on its own. But other controls lean on it. An agent missing from the inventory is likely one whose owner is hard to name quickly.
Produced by the WebPulse Newsroom with AI assistance from the original reporting credited below, and checked against that source by our editorial review. How we use AI.
Original reporting: Wiz.
