"Agents Without Guardrails": 65% of Enterprises Have Watched an AI Agent Act Outside Its Intended Scope

202 enterprise technology and security leaders surveyed by Enterprise Management Associates for Cequence Security

65% of respondents reported observing AI agents operating beyond their intended scope, and 29% reported measurable organizational impact from it. A further 35.6% caught near-misses before material harm, and 3.5%, seven organizations, first learned of an incident from a customer or partner rather than from their own monitoring. On containment, 32.2% could detect and contain an out-of-scope action within minutes using automation while 54.5% needed hours and manual intervention, and 46% could not readily produce a complete 30-day audit trail of agent activity. On authorization, only 34.2% evaluated agent authorization at execution time; 94% were confident their agents held no excess access, yet only 32.7% provisioned agents with least privilege, and 47% lacked a reliable agent inventory. 54.5% required and enforced unique agent identities and another 32.2% required them without consistent enforcement. Deployment is not slowing on paper (46% were scaling agentic AI across multiple departments and production workflows, 79% ran generative and agentic AI in parallel) but 30% of agentic pilots had been paused or discontinued, with 48.5% citing security risk

Not applicable. The gap the survey measures is between confidence and instrumentation: near-universal belief that agents are correctly scoped, a minority actually provisioning least privilege, and a majority unable to evaluate authorization at the moment an action is taken