Israeli Expert Warns AI Dependence Poses Strategic Risks Without Governance

Dr. Maoz Rosenthal warns that Israel's reliance on foreign artificial intelligence models could become a strategic vulnerability. He stresses the urgent need for governance and safety mechanisms to ensure operational continuity during crises. Rosenthal cites the June 2026 incident when Anthropic released Claude Fable 5, its most powerful AI model to date, which was quickly found to have security flaws. The U.S. government immediately blocked all foreign access, including from foreign employees, forcing Anthropic to shut down the model entirely. This event highlights the security risks of advanced AI and the fragility of sovereignty for countries like Israel that depend on American AI models, which can be cut off unilaterally.

Rosenthal criticizes Israel's contradictory approach: it seeks minimal regulation to foster a thriving AI startup ecosystem while expecting state protection during crises, such as defense against hostile algorithmic weapons or ensuring access to AI models amid geopolitical conflicts. He draws parallels to the NSO Group Pegasus scandal, where Israel allowed lax regulation to promote innovation but ultimately bore the diplomatic fallout when the company faced U.S. sanctions for selling hacking tools to authoritarian regimes.

He warns that many Israeli public institutions, including security agencies and courts, rely on AI infrastructure not owned or controlled by Israel, creating structural dependence. Technology companies resist binding regulations to maintain growth but expect government protection when risks materialize, creating a systemic paradox. Rosenthal argues Israel cannot sustain this duality without paying a price.

To address this, he proposes establishing a central governmental AI safety lab with executive authority to rigorously test AI models before deployment in public services and maintain a sensitive data layer fully owned by Israel, isolated from foreign providers. This would mitigate risks if access to external AI models is cut off during crises. While not a complete solution to dependence on global computing infrastructure, it is a necessary first step to ensure AI models used domestically cannot be exploited against Israel.

Rosenthal concludes that proactive caution was reasonable when AI was experimental but is now reckless as AI systems manage critical functions like legal evidence analysis and national security decisions. Israel must build institutional infrastructure to guarantee both innovation and sovereignty, rather than relying solely on market forces to provide them.