AI Security Survey Finds Major Gaps in Governance, Visibility and Agent Controls

Summary

A Cybersecurity Insiders survey of 1,253 professionals, supported by Netskope, finds widespread AI security gaps: 94% report incomplete visibility, and only 9% say they can prevent risky agent actions before execution.

Key points

The report surveyed 1,253 cybersecurity and IT professionals in early 2026 using stratified sampling; it states a 95% confidence level and a ±2.8% margin of error.
AI tools are deployed at 73% of surveyed organizations, while only 7% report advanced governance with real-time policy enforcement.
Only 6% report complete visibility into AI activity, and only 12% can reliably distinguish personal from corporate AI accounts.
Only 8% say their controls detect sensitive meaning after AI rewrites content; 46% say rewriting usually bypasses their regex or keyword matching.
AI tools or agents have write access to collaboration systems at 53% of organizations, code repositories at 25%, and identity providers at 8%.
Only 9% report full pre-execution prevention for risky agent actions. Separately, 37% report AI-caused operational issues during the past year, including 8% reporting significant disruption.
The report recommends broader monitoring of API and machine-to-machine activity, semantic data inspection, approval gates for high-risk agent actions, and stronger non-human identity governance.

Article Details

Publisher: Cybersecurity Insiders
Report Period: Survey conducted in early 2026
Scope: Enterprise AI security governance, visibility, data protection, non-human identity management, and autonomous-agent controls across industries and organization sizes.
Sample Size: 1,253 cybersecurity and IT professionals; stratified sampling with a reported 95% confidence level and ±2.8% margin of error.
Key Statistics: 73% of surveyed organizations deploy AI tools, while only 7% report advanced governance with real-time policy enforcement, a 66-percentage-point gap.
94% report incomplete visibility into AI usage; 88% cannot reliably distinguish personal AI accounts from corporate instances.
Only 8% confirm semantic detection of sensitive content after AI rewriting; 46% say rewriting typically bypasses their regex or keyword controls.
Only 9% report inline prevention for risky API or agent actions; 24% can prevent some actions, 35% would discover them only afterward in logs, and 32% have no visibility into agent actions.
37% report AI-caused operational issues in the preceding year, including 8% reporting significant disruption such as outages or data corruption.
Recommendations: Expand activity-level monitoring across SaaS, API, and machine-to-machine traffic, distinguishing personal from corporate AI accounts.
Identify the three highest-risk AI use cases, embed enforceable policies in technical controls, and assign an owner to each.
Test whether DLP detects sensitive information after AI rewriting and use the results to guide semantic inspection at the point of transfer.
Audit agent write permissions, require approval for account creation, permission changes, and external data transfers, and establish agent-specific detection and request-layer containment.
Unify CASB, DLP, and access-policy evaluation so content classification, identity, and AI instance type inform enforcement before execution.