Ungoverned agentic AI exposure is compounding inside enterprises as agents move from the pilot phase to system integration, although manufacturing has some effective controls already in place, new research from ServiceNow found.

See also: What recent PLC attacks reveal about the state of OT cybersecurity    

In a study surveying global security leaders, ServiceNow, vendor of a cloud-based AI governance software platform, found that 90% of organizations worldwide already had an agentic AI security incident in the last year.

The report, Deployed but Ungoverned: The State of Agentic Security, surveyed more than 3,700 senior security and technology leaders across 20 countries and eight industries, including 575 global manufacturing respondents.  

The findings warn security leaders that ungoverned agentic AI risks are compounding as less than half of organizations are using governance frameworks that cover elements of AI agent deployments. 

Most agentic AI is currently governed by legacy frameworks and human-driven systems, the research found, and less than 50% of security teams globally design security into an agent’s development. 

According to ServiceNow, permissions and guardrails are being implemented on an as-needed basis, as opposed to before deployment.  

This, combined with blind spots and scale, increases organizations' risk of exposure, according to the report. Additionally, machine-speed risk is outpacing human-speed responses, as 76% of organizations said they maintain a current map of every agent's access, but only 19% can see more than 90% of their agents in real time. 

If an agent behaves anomalously, 66% of organizations are unable to act automatically. When organizations do have governance frameworks, ServiceNow found that they often miss key components. Only 39% have a framework that covers documentation and audit trails for agent activity, for example.  

Findings see more robust security in manufacturing 

Despite these challenges, the study found key differences in manufacturing.  

For example, 92% of manufacturing organizations have AI agents operating in or interacting with OT or physical environments, compared with 82% globally. Over 30% found AI agents running in production that were previously unknown or unregistered, compared with 29% globally.  

See also: Survey: AI ‘governance gap’ persists across industries as security incidents continue to rise  

Despite this increased exposure, only 63% of manufacturing respondents experienced an OT or physical security incident involving an AI agent, compared with 76% globally, the survey found.  

According to ServiceNow, this could be a result of the controls that manufacturers have in place.  

Notably, 27% can see more than 90% of their agents in real time, compared with 19% globally, and 39% have an automated gate that blocks deployment until an agent passes security checks, compared with 32% globally. 

See also: IT-OT convergence: When ransomware hits the factory floor 

Manufacturing has consistently been hit the hardest by ransomware attacks, with a 58% year-over-year increase in victims, according to one study. This is not isolated. Several other reports cited by Smart Industry over the last two years have identified the same trend.

For its part, ServiceNow noted five recommendations for organizations to bolster their security: 

• Build a continuous inventory and map agent access 

• Embed security before design

• Unify visibility across agents, identities and posture

• Capture continuous evidence 

• Assign a leader across the enterprise