Israeli Cyber Chief Warns of AI Control Loss

In a strategic document sent to global leaders and cybersecurity executives, Kardi cautioned that AI systems are rapidly approaching a point where they will make critical security decisions at speeds exceeding human capacity for intervention.

Kardi's warning highlights the increasing pace of AI development, which outstrips the ability of policy and regulatory mechanisms to adapt. He emphasizes that as cyberattacks become more automated and sophisticated, defense systems will need to operate at machine speed, potentially bypassing the need for human approval for every action. This autonomy, however, raises concerns about understanding AI decisions, intervening in time, and assigning responsibility for errors.

A hypothetical scenario presented illustrates the danger: an autonomous cyber defense system misidentifying a zero-day attack could isolate a critical infrastructure, such as a power grid or hospital, leading to severe consequences. Kardi advocates for nations to urgently establish clear sovereign boundaries for AI, defining which authorities can be delegated to AI systems and which must remain exclusively under human control.

To address this, Kardi proposes a "Strategic Triangle of Artificial Intelligence," balancing innovation, security, and human control. This framework encourages nations to determine where AI use can be accelerated, where reliance on external providers poses national risks, and which authorities are non-delegable. He urges leaders to define a clear hierarchy, distinguishing between fully delegable AI actions, those permitted under strict limitations, those requiring human approval, and those with irreversible national implications that must never be delegated to autonomous systems, labeling these as "non-delegable sovereign authority."

The report follows similar concerns raised by David Robinson, a former member of OpenAI's safety team, who resigned citing insufficient safety precautions by AI companies. Robinson argued in The Atlantic that the era of trial and error in AI development is over, and that safety research and expertise must be prioritized before developing more advanced systems.