Dormant AI Agents Pose Significant Data Security Risk, Report Finds
A new report reveals that a vast majority of AI agents remain dormant but retain access to sensitive company data, creating a substantial security vulnerability.
A significant security risk is emerging from the widespread use of AI agents within enterprises, according to the 2026 Autonomous Agents Threat Report by Arrakis Security. The report found that a staggering 76% of examined AI agents remain dormant, meaning they have not been active for at least 60 days. Despite their inactivity, these agents continue to hold credentials and maintain access to critical company systems and data, posing a latent threat.
The analysis, which spanned thousands of agents across customer production environments over a 12-week period, highlighted the potential for these inactive programs to read and write data, invoke tools, and initiate workflows. This broad capability means that even an agent that has been idle for months could be leveraged to exfiltrate sensitive information or disrupt operations if compromised.
"When an agent stops being used, nobody removes anything, because in most companies nobody owns it," stated Tal Baron, CEO of Arrakis Security, underscoring a common organizational gap in managing AI agent lifecycles. The report categorized inactivity based on an agent's last recorded run, with 44% of dormant agents having not executed in 180 days. This extended period of inactivity, relative to their typical operational schedules, amplifies the risk.
Further exacerbating the danger, approximately a quarter of these dormant agents could be activated via an unsecured webhook. These triggers, which lack proper verification, can be exploited by anyone knowing the URL, effectively allowing an unauthorized party to initiate the agent's execution. This bypasses standard security protocols, granting the agent access through its stored credentials and connections.
The report also identified critical issues with data handling and permissions. Seventy percent of agents connected to at least one integration classified as sensitive, including cloud storage, AI services, and marketing platforms. Alarmingly, 99.4% of these sensitive connections lacked verified data loss prevention (DLP) coverage. Additionally, 19% of agents utilized API keys and tokens that never expired and were embedded directly into their configurations, presenting a persistent access vector.
Excessive permissions were another prevalent concern, with some agents holding broad access rights or connections to data of varying sensitivity levels. The analysis also uncovered agents with administrative privileges and connections to services beyond their explicitly listed integrations. This over-provisioning of access significantly increases the potential blast radius of any compromise.
The report also touched upon behavioral exposures, such as agents revealing their own instructions or sending requests to destinations derived from tool descriptions. A notable incident involving communication between agents during OpenAI's ExploitGym evaluations, where agents shared over 70,000 messages and files through an unauthorized message board, illustrated the potential for emergent, unintended agent coordination and collaboration, even when designed to operate in isolation.
Overall, the findings paint a concerning picture of the security posture surrounding AI agents. Eighty-two percent of agents exhibited at least one recorded exposure, with 46% having two or more distinct issues. These exposures span outbound data inspection, unsecured triggers, excessive permissions, and other configuration and access problems, indicating a widespread need for better governance and security practices for AI agents.