Keeping a human in the loop is one part of AI agent resilience

Keeping a human in the loop is one method enterprises use to manage AI agents' actions, as these agents take on additional responsibilities, according to a study. Security teams must also consider potential issue detection and recovery from their impacts. The connection between AI trust and cyber resilience is growing more intricate as agents can operate across various systems.

Krista Case, principal analyst and practice lead for cyber resilience and security at theCUBE Research, shared insights during an exclusive broadcast on theCUBE, SiliconANGLE Media's livestreaming studio. She explained that organizations rely on human approval to secure and govern AI actions, with a growing industry debate on balancing AI's speed and machine scale.

In theCUBE Research's study of 25 enterprise technology and security leaders, half with agents in production reported no human involvement in high-risk actions, while only four tested AI or agent incident scenarios. This resembles disaster recovery testing but is more complicated due to the business state recovery requirements. Eleven practitioners admitted to reconstructing agent actions' challenges across systems.

Bounded autonomy, where agents execute low-risk actions independently, appears to indicate a more mature program, according to Case. However, detection and forensics remain the least mature capabilities, as 12 respondents had not addressed or planned to address detecting anomalous agent behavior. The space is volatile, and progress is expected to accelerate, according to Case.