Enterprises Grapple with Unseen AI Agent Actions, Survey Finds
A new survey reveals that nearly half of organizations struggle to track, control, or automatically halt the actions of AI agents, leading to concerns about unauthorized data modification, permission changes, and external API calls.
Enterprises are rapidly adopting AI agents, but a significant gap exists between their deployment and the ability to govern their operations, according to a recent survey by Salt Security. The "AI and API Security Trends 2026 (H2)" report indicates that 54% of organizations suspect an AI agent has taken an unintended, unexpected, or unauthorized action within the past year. These actions range from external API calls and transaction anomalies to unauthorized configuration changes and data modifications or deletions.
The rapid integration of AI agents into business processes is outpacing the development of robust governance and runtime controls. This disconnect leaves organizations vulnerable to unforeseen consequences. "The data shows a gap that is already producing real operational consequences," stated Roey Eliyahu, CEO of Salt Security. "Organizations are deploying agents and granting them access to enterprise data, APIs, and services before they can trace what those agents are doing or stop them automatically when something goes wrong."
The survey, which polled over 380 security, technology, and DevOps professionals, found that AI agent usage is widespread, with 87% of respondents reporting at least one agent operating within their organization, and 60% managing more than ten agents. A common practice involves agents invoking other agents or tools, which then call APIs, creating complex chains of actions across multiple systems. This interconnectedness makes it exceedingly difficult to trace an agent's initial instruction to its ultimate outcome.
Visibility into these complex agent interactions is severely limited. Only 12.5% of respondents reported being able to consistently trace the full sequence of an agent's activity from instruction to outcome across all systems. Furthermore, the ability to intervene when an agent behaves unexpectedly is also a significant challenge. Merely 17.5% of organizations can automatically halt a rogue agent in production, with many resorting to manual intervention that can take minutes or even require taking an application offline.
Identity and permission management for AI agents also presents a hurdle. While some organizations assign dedicated identities with scoped permissions (26%), a substantial portion uses dedicated identities with broad permissions (24%), or has varied approaches depending on the application. Alarmingly, over 17% admitted they didn't know how agent access was assigned, highlighting a critical security blind spot.
Beyond agent actions, the report also flagged concerns about "shadow MCP infrastructure" - unauthorized, shadow, or misconfigured management and control plane servers. Salt Labs telemetry indicated a roughly 30% relative increase in the detection of such infrastructure between January and August 2026.
To address these growing challenges, Salt Security recommends establishing comprehensive agent governance systems. Key strategies include maintaining an up-to-date inventory of all agents, tools, and APIs; assigning agents strictly limited permissions; designating accountable owners for each agent; and implementing pre-approval checks for high-impact actions before they are executed against business systems.