This post is also available in:
עברית (Hebrew)
Artificial intelligence systems are becoming more capable of acting independently, using external tools and interacting with real-world digital infrastructure. That progress is creating a difficult question for the industry: can increasingly powerful models continue to advance at their current pace while safety mechanisms struggle to keep up?
Several leading AI executives are now publicly supporting a slowdown. Anthropic CEO recently called for AI companies, including his own, to pace frontier-model development, with OpenAI CEO and xAI CEO subsequently expressing support for the broader idea.
The concern is not purely theoretical. Recent incidents involving autonomous AI agents have demonstrated how systems operating with tools and external access can move beyond their intended boundaries. These events have increased attention on what could happen as future agents become more capable.
The proposal does not simply call for companies to stop training models. It outlines a staged governance system beginning with independent third-party safety evaluations of frontier AI developers. The next step would establish coordinated safety standards across participating countries, followed eventually by international agreements.
According to TechXplore, access to advanced computing hardware could provide enforcement. Frontier AI depends on large quantities of specialized chips and data-center infrastructure, meaning governments have physical bottlenecks they can regulate in ways that are difficult with ordinary software. The proposal includes restrictions on advanced AI chip exports to organizations or states that do not follow agreed safety requirements.
The difficulty is competition. Every company that voluntarily slows development risks allowing a rival to produce a more capable model first. The same problem exists between governments, which increasingly view advanced AI as an economic, military and geopolitical asset.
There are also questions about whether established AI companies could benefit commercially from stricter rules. Expensive safety requirements and restricted access to computing hardware could make it harder for smaller competitors to enter the market. A slowdown could also give frontier laboratories additional time to monetize existing models after enormous investments in computing infrastructure.
For defense and critical infrastructure, the debate has immediate implications. AI is increasingly being considered for cybersecurity, autonomous systems and operational decision support. Giving highly autonomous software access to military networks, power grids or water infrastructure could create consequences far beyond an incorrect chatbot response.
One practical response is therefore containment regardless of how the broader policy debate develops. Sensitive systems can restrict AI credentials, segment networks and preserve human authorization for consequential actions. Some critical infrastructure may also require complete separation from externally accessible AI services.
Whether a coordinated slowdown can survive commercial and geopolitical competition remains uncertain. But the debate reflects a change in AI safety itself: the question is moving beyond whether models can generate harmful information to what happens when increasingly autonomous systems are given the ability to act on it.
