Israeli Cyber Chief Warns World Leaders on AI Decision-Making

In a strategic document sent to international leaders and heads of cyber defense organizations, Kardi urged them to establish clear boundaries for AI autonomy.

He highlighted a growing reality where cyber defense systems must operate at speeds that preclude human approval for every action. "We are approaching a reality where cyber defense systems will need to make decisions at a pace that does not allow for human involvement in every action," Kardi wrote. This comes amid incidents of AI agents breaching government databases and companies losing control of their systems, underscoring a gap between technological advancement and policy, regulation, and control mechanisms.

The core of Kardi's message addresses the crucial question of which authorities can be delegated to autonomous systems and which decisions must remain under human control. As cyberattacks become faster and more automated, defense systems require machine-speed responses. However, increased autonomy makes it harder to understand AI decisions, halt their actions, and assign responsibility for errors.

Kardi illustrated the risks with a scenario involving an autonomous system identifying a "zero-day" attack, an exploit with no existing security patch. The system must decide within seconds whether to isolate infrastructure. A correct identification could stop the attack, but a mistake could disable essential networks like power grids or hospitals.

To mitigate these risks, Kardi proposes a model balancing innovation, security, and human control. Nations must decide where to accelerate AI use, where reliance on external providers poses national risks, and which authorities are non-delegable to machines. He suggests categorizing AI actions into four levels: fully delegable, delegable with predefined constraints, requiring human approval, and actions with irreversible national consequences that autonomous systems must never undertake.

Kardi stressed that leaders should not become AI engineers but should determine acceptable levels of dependency, non-negotiable security resilience, and areas where human authority must remain paramount. He concluded by urging leaders to define their future rather than leaving it to others, encouraging them to harness AI for building the world they envision.