Agentic AI,
Artificial Intelligence & Machine Learning,
Next-Generation Technologies & Secure Development
Agent Governance and Runtime Controls Have Yet to Catch Up With Adoption
Recent research indicates a significant challenge faced by organizations in managing artificial intelligence (AI) agents. The study revealed that nearly half of the organizations surveyed suspect or confirm that an AI agent operated outside their intentions in the past year. This data presents a growing concern for various business sectors as AI technology continues to proliferate.
The “AI and API Security Trends 2026 (H2)” report, published by cybersecurity firm Salt Security, highlighted that 54% of respondents reported instances involving external API calls or transactions initiated by AI agents. Along with that, half of the respondents noted configuration or permission changes involving these agents, whereas 42% indicated occurrences of data being modified or deleted without proper authorization. Notably, participants were allowed to select more than one outcome, emphasizing the multiplicity of issues that organizations face.
Roey Eliyahu, CEO of Salt Security, addressed this alarming trend in a statement, stating, “The data shows a gap that is already producing real operational consequences.” He emphasized that many organizations are deploying AI agents rapidly, granting them access to crucial enterprise data, APIs, and services before establishing the necessary controls to monitor their actions effectively. This gap in governance is leading to unintended operational mishaps that could have been avoided with better oversight.
The report draws on insights from over 380 professionals working in security, technology, architecture, application security, API platforms, and DevOps roles. Furthermore, the report incorporates telemetry data from Salt Labs monitoring hundreds of customer environments, providing a robust overview of the current state of AI agent governance.
Widespread usage of AI agents was noted among respondents, with 87% revealing that at least one AI agent operates within their organization. More strikingly, 60% of these organizations reported having more than ten AI agents in play. The report further indicates that over three-quarters of respondents regularly observed their agents invoking other agents or tools that, in turn, call APIs, resulting in a chain reaction of actions taken across various systems.
Given this interconnected nature, tracking an agent’s initial instruction through to its resultant actions becomes problematic. Alarmingly, only 12.5% of respondents claimed they could consistently trace this entire process across all AI agents in their organization, highlighting a significant oversight in operational capabilities.
In a query about reconstructing an AI agent’s activity from the previous day, only 15.7% of survey takers indicated that they could perform this task within minutes using automated tools. The majority, comprising about two-thirds, said they would require hours or even days of manual effort to trace back an agent’s activity, a considerable resource drain.
Moreover, stopping an agent mid-performance poses an additional challenge. A mere 17.5% of organizations affirmed that they could automatically interrupt an agent that exhibited unexpected behaviors during production. The rest of the respondents mentioned varying degrees of manual intervention, with over half stating they could halt an agent manually within minutes. In contrast, 12% had to resort to taking the entire application offline to stop an agent’s action effectively.
The survey also highlighted how organizations manage agent identities and permissions. Only 26% of respondents indicated that agents receive dedicated identities with specifically scoped permissions. Meanwhile, around 24% reported that their agents operated with broader permissions under dedicated identities, and over 17% varied permissions based on the application context. Alarmingly, about 11% expressed uncertainty regarding how agent access was allocated.
Furthermore, operational issues related to shadow Managed Cloud Platform (MCP) infrastructure were noted, with 33% of respondents having discovered an unauthorized, misconfigured, or shadow MCP server within the past year. This finding is complemented by Salt Labs’ discovery that 21.6% of customer environments contained MCP-related infrastructure in August 2026, a noticeable increase from 16.7% in January, representing a roughly 30% rise.
To rectify these governance challenges surrounding AI agents, Salt Security proposes several measures. These include maintaining an inventory of all agents, MCPS, tools, and APIs, assigning agents limited permissions, designating accountable owners, and conducting a thorough check of high-impact actions before they reach essential business systems. By implementing these strategies, organizations can enhance their governance frameworks and minimize unforeseen operational disruptions caused by AI agents.