Translated from Korean

If it were up to me, I would attach specific execution permissions next to the 'urgency' labels provided by security AI. There is more to verify between deciding that an issue needs quick attention and deciding that a service should be shut down immediately.

CISA's SSVC prioritizes responses based on factors like actual exploitation, technical impact, and the potential effect on organizational mission and people. Even the most urgent 'Act' category is described as a process where supervisors and responsible parties define and agree upon necessary actions. This label itself does not grant AI the authority to perform automatic shutdowns.

You also need to look at product execution settings separately. Microsoft Defender documentation distinguishes between automated actions and pending approval based on device group settings, allowing users to review, approve, deny, or check history in the Action Center. While some completed actions can be undone, the support scope is limited. This is a description of that specific product's automated investigation and remediation, and it does not mean all generative AI supports the same functions.

Based on these two resources, I believe it would be beneficial for organizations to review and divide permissions as follows. Below is a suggested example.

• Automate the process of reading logs for defined targets and grouping the findings for notification, but separate the authority required to make actual changes. Include the reasoning and any remaining uncertainties with the alerts.

• For blocking or isolation of restricted targets, define the targets, scope, duration, release conditions, and responsible personnel in advance. Also, verify whether an action is reversible through actual recovery procedures rather than just the feature name.

• Require additional approval for blocks affecting multiple services or changes that are difficult to reverse. It would be helpful to show the expected impact on normal operations and recovery methods on the approval screen.

However, waiting for human response for every action could lead to delays during an attack. If there is a need for emergency exceptions, you must define in advance the conditions, scope, when to escalate to humans, and when to halt. This example is not a one-size-fits-all answer for every organization. Since Microsoft also recommends full automation for its specific product, you should check actual product features and your organization's operational readiness together.

Furthermore, even if an action can be undone, the damage caused by an interrupted service is not erased. It is recommended to track metrics such as the number of false-positive blocks, the impact on normal operations, approval wait times, and service recovery times. This will serve as a basis to evaluate whether the AI is accurately identifying issues and whether the granted permissions were appropriate.

Loading comments…