Israel's cyber chief urges world leaders to set limits on AI authority

The head of Israel's national cyber agency has written to world leaders and the heads of foreign cyber defence organisations, warning that governments risk losing control over artificial intelligence agents and models. According to an Israeli government press release, Yossi Karadi, Director General of the Israel National Cyber Directorate, set out his concerns in a strategic paper and called on nations to define, without delay, the limits of the authority they hand to machines.

The central argument, as the release summarises it, is that cyber defence is moving toward a point where systems must decide and act at a speed that leaves no room for a person to approve each step. Karadi asks national leaders to settle now which powers can be passed to AI systems and which must stay with people alone.

Who is sending the warning

The Israel National Cyber Directorate, abbreviated INCD, is the Israeli government body responsible for national cyber defence. Karadi heads it with the title of Director General. The paper was addressed both to political leaders and to the chiefs of cyber defence organisations in other countries, making it an appeal to counterparts abroad rather than a domestic policy announcement.

The release links the timing of the paper to recent cases in various countries where AI agents broke into government databases, and to many further cases in which companies could no longer control what their own agents were doing. The release does not identify any of these incidents.

The problem of machine speed

Karadi's paper describes cyberattacks becoming quicker, more automated and more precisely aimed. In response, defensive systems will also have to spot threats, reach decisions and respond at machine pace, beyond what human sign-off can keep up with. In his words, cyber defence is heading toward decisions made "at a pace that does not allow human involvement" in each action.

The paper identifies a tension in this shift. Giving a system more independence makes its choices harder to understand, makes it harder to halt in time, and blurs who is responsible if it gets something wrong. Karadi also writes that AI is developing faster than states can update their policy, regulation and oversight.

To illustrate the danger, the paper offers a hypothetical. An autonomous defence system detects an attack using a "zero day" vulnerability, meaning a flaw for which no security fix yet exists, and must cut off parts of the infrastructure within seconds. If its judgement is right, the attack could be stopped. If the detection is false, the system could shut down critical services, with a power grid or a hospital given as examples.

The proposed framework

Karadi's answer is what the paper calls a strategic AI triangle, published as a document titled The AIST AI Strategic Triangle. It asks governments to weigh three interests against each other: innovation, security and human control.

Under this approach, the release says, adopting AI is taken as given. The questions become where its use should be sped up, where relying on outside suppliers and infrastructure poses a national risk, and which powers should never be handed to machines.

The paper then proposes that countries sort decisions into a ladder of four categories:

• actions AI may carry out entirely on its own;
• actions AI may take only inside limits set in advance;
• actions that need a human to approve them;
• actions with irreversible national consequences that no autonomous system should ever take, which the paper labels "non delegable sovereign authority."

Karadi frames the task as one for leaders rather than technicians. "Leaders do not need to become AI engineers," he writes, adding that their job is to judge where dependence is tolerable, where resilience cannot be compromised and where human authority must prevail. The paper closes with an exhortation: "Do not leave it to others to define your future."

What the reports do not say

This account rests on a single source, the Israeli government's own press release, so the description of the paper's content is the directorate's presentation of its own work. No outside organisation's coverage or reaction was available for this report.

The release does not name the world leaders or foreign cyber organisations that received the paper, or say how many there were. It does not identify the incidents involving AI agents breaching government databases, the countries where they occurred, or the companies said to have lost control of their agents.

It also does not say whether Israel itself has adopted the four-tier hierarchy in its own cyber defence, whether the proposal is tied to any international forum or negotiation, or whether any government has responded. The release refers readers to the full paper for further detail, and the summary here does not draw on anything beyond what the release itself states.