Enterprises are accelerating the deployment of autonomous AI agents across their operations while falling dangerously behind on fundamental governance and security controls. This is the central conclusion of the third State of AI Agent Security Report published by enterprise agent management firm Gravitee, which surveyed 900 IT and technology executives across the United States and the United Kingdom. Within six months, the share of organizations running more than 150 autonomous agents in parallel more than doubled, climbing from 7 percent to 16 percent. These software entities increasingly execute complex tasks, access production databases, and make automated operational decisions.
This rapid scaling stands in stark contrast to the existing governance infrastructure within enterprise IT environments. Only 17 percent of surveyed organizations assign distinct, isolated identities or credentials to their autonomous agents. The deficit in human accountability is even more severe: fewer than 10 percent of companies have established a named individual who is formally liable for the actions, errors, or financial consequences of a running agent. Furthermore, enforceable behavioral policies currently exist for only 51 percent of active enterprise agents.
The real-world consequences of this accountability vacuum are already materializing across industries. A majority of respondents, exactly 52 percent, reported experiencing a security or data breach over the past twelve months that was directly triggered by an autonomous AI agent. These incidents span unintended data disclosures, unauthorized system modifications, and execution anomalies that cascade across interconnected enterprise tools. Operating without continuous human oversight, agents frequently trigger critical exceptions that traditional monitoring stacks fail to isolate.
As operational incidents mount, executive confidence in managing agent-specific risks is deteriorating rapidly. While 47 percent of technology leaders considered their organizations very well prepared to handle agent-related incidents six months ago, that figure has plummeted to just 29 percent in the latest survey. Many organizations are realizing after deployment that legacy security postures, designed for static software and user accounts, are poorly equipped to govern non-human actors with dynamic execution capabilities.
Sustaining agent operations will require enterprises to establish dedicated identity frameworks, strict permission boundaries, and reliable rollback capabilities. When autonomous pipelines operate without robust approval gates, organizations face substantial operational drag and post-incident remediation costs to clean corrupted data. Companies deploying autonomous software into production environments must establish clear ownership and technical access controls before granting write permissions to core business systems.