Survey: enterprises rarely test AI agent incident scenarios

A theCUBE Research study of 25 enterprise tech and security leaders found that none of the 19 organizations running AI agents in production allow high-risk actions without human approval, yet only four of 25 had tested an AI agent incident scenario. Eleven of 25 acknowledged gaps in reconstructing agent actions across systems.

• 19 of 25 respondents already run AI agents in production
• Only 4 of 25 tested an AI agent incident scenario
• 11 of 25 cannot reconstruct the sequence of agent actions
• 12 of 25 are not addressing detection of anomalous agent behavior