News
Companies Struggle to Track and Stop AI Agents, Study Finds
Companies are deploying AI agents faster than they can monitor and control them effectively, according to a new report by Salt Security.
The company's "State of Agentic AI and API Security, 2H 2026" report found that 49.8 percent of respondents confirmed or suspected that an AI agent took an action their organization did not intend, expect, or explicitly authorize in the past 12 months.
The findings are based on a survey of more than 380 security and technology professionals, alongside production API security data collected by Salt Labs.
AI Agents Are Taking Unexpected Actions
Of those reporting confirmed or suspected unintended actions, 53.9 percent said the incidents involved external API calls or transactions, while 49.7 percent reported changes to configurations or permissions.
Another 41.9 percent said agents modified or deleted data, and 32.5 percent reported agents accessing information they should not have been able to reach.
However, organizations also reported limited visibility into what their AI agents were doing.
Only 12.5 percent said they could consistently trace the complete sequence of actions across all agents, from initial instructions through tools and APIs to the outcome.
Another 38.6 percent could do so for most agents, while 25.6 percent had that visibility only for selected systems.
The research also found that just 15.7 percent could reconstruct everything an AI agent had done the previous day within minutes using automated tools.
By comparison, 36 percent would require hours of manual investigation, while 28.5 percent would need days of work involving multiple teams.
Most Companies Cannot Automatically Stop AI Agents
The report identified similar limitations in organizations' ability to intervene when agents behave unexpectedly.
Only 17.5 percent of respondents said their organizations could automatically stop an AI agent mid-action in a production environment.
Another 56.1 percent could intervene manually within minutes, while 12 percent would need to shut down the application. A further 3.1 percent said they could not stop the agent.
Meanwhile, 86.4 percent reported having at least one AI agent operating within their organization, and 59.5 percent had more than 10.
Salt Security recommended that organizations maintain inventories of their AI agents, assign dedicated identities with limited permissions, and implement monitoring that tracks actions across connected tools and APIs.
The report also called for automated controls that can block high-risk actions before they are executed.
“The defining security change in the agentic era is not simply that AI can generate more content or code. It is that AI systems are increasingly authorized to act,” the report stated.