Surveys Highlight Gaps in Enterprise AI Agent Security and Governance

Summary

Surveys cited by Gravitee report limited visibility and access controls for enterprise AI agents, with organizations reporting unauthorized behavior and security incidents. The article recommends agent inventories, scoped authorization, monitoring, and audit trails.

Key points

A Cybersecurity Insiders and Saviynt survey of 235 CISOs at organizations with 5,000 or more employees found 71% had AI agents with direct access to core systems, while 16% effectively governed that access.
The article says 92% of surveyed organizations lacked full visibility into AI identities, 95% doubted they could detect or contain misuse, and 47% had observed unintended or unauthorized agent behavior.
Gravitee says its survey of 919 executives and practitioners found 88% of organizations confirmed or suspected security incidents involving AI agents that year.
The article reports that 75% of organizations had discovered unsanctioned AI tools, while only 14.4% had full IT and security approval for their agent fleets.
Recommended measures include inventorying sanctioned and unsanctioned agents, assigning each an identity and owner, limiting token scope, and using contextual authorization.
The article also recommends continuous monitoring and auditable access decisions; it argues that cumbersome approved tools can encourage shadow AI use.

Article Details

Publisher: Gravitee
Report Period: 2026
Scope: Enterprise AI agent security and governance, including visibility, identity, authorization, monitoring, and regulatory considerations.
Sample Size: Gravitee surveyed 919 executives and practitioners; Cybersecurity Insiders surveyed 235 CISOs at organizations with 5,000+ employees.
Key Statistics: 71% of large enterprises reportedly deployed AI agents with direct access to core business systems; only 16% effectively governed that access.
92% of organizations reportedly lacked full visibility into AI identities, while 95% doubted they could detect or contain misuse.
86% of organizations reportedly did not enforce access policies for AI identities.
75% of organizations reportedly had discovered unsanctioned AI tools in their environments.
Only 5% of organizations reportedly felt prepared to contain a compromised AI agent.
Recommendations: Inventory sanctioned and unsanctioned AI identities and agents, including their access footprints, embedded credentials, service accounts, and delegated agents.
Establish ownership and lifecycle management for each AI agent, with explicit permissions and access expiration or renewal.
Use fine-grained, contextual authorization for AI identities rather than relying only on coarse-grained role-based access control.
Use scope-limited tokens bound to specific resources, operations, and time windows.
Continuously monitor agent activity and permissions, and maintain auditable records of authorization decisions.