[Wafer: response was truncated before the model finished its internal reasoning. Increase max_tokens, or disable thinking on this model (e.g. chat_template_kwargs.enable_thinking=false), then retry.]

The Trump administration imposed sweeping restrictions on the most advanced AI models from Anthropic and OpenAI in June 2026, halting their rollout and limiting access to a small group of government-approved US companies. The reversal came after US intelligence agencies demonstrated that safeguards in the models could be bypassed, raising fears that rivals could use them to attack critical infrastructure.

From deregulation to restrictions

President Donald Trump had spent much of his term easing regulations to accelerate AI adoption across government and industry. The US military used AI models, including specialized versions, during the war with Iran to process intelligence and help select strike targets. The restrictions on Anthropic and OpenAI followed an alarming leap in the models' cyber capabilities. Intelligence agencies feared that China, Russia, or hostile hackers could use them to disrupt or cripple critical national infrastructure. The companies said they had built safeguards to prevent the models from identifying and exploiting security vulnerabilities, including limits on generating attack code. US intelligence agencies argued and demonstrated that those safeguards could be bypassed, which prompted the White House to act. Previous major cyberattacks caused extensive economic damage, but cybersecurity teams typically contained them within days. Malware generated by AI rather than a human operator could prove far harder to stop. Under pressure, Anthropic and OpenAI complied. By the end of June 2026, the models' rollout had been halted, and they are now available only to authorized American organizations working to protect critical infrastructure.

China's control-based system

Beijing has also built a complex relationship with AI, driven by two competing goals: matching US technological leadership and maintaining strict control over information and public opinion. China's regulatory system is among the fastest-developing in the world. Official rules require AI-generated content to "reflect the core socialist values," meaning models cannot produce content critical of the Communist Party leadership or President Xi Jinping, or present alternative interpretations of historical events such as the Tiananmen Square crackdown. Companies must also train their models on "screened and approved" information to prevent them from absorbing Western democratic ideas. The system works in reverse of the US approach. Algorithms must receive advance approval, be registered in a government database, and pass safety testing before release. If a model produces an answer that violates government requirements, the developing company faces legal and criminal liability. In practice, Chinese firms impose extensive self-censorship on their systems. At the same time, Beijing was among the first governments to introduce detailed rules targeting deepfake tools. AI-generated images and videos must carry labels indicating they were artificially produced, and using such technology to imitate a person's voice or appearance without consent is broadly restricted. Some experts say China's limits on training data and model freedom help explain why the United States still holds an edge in AI capability. That balance may shift with the release of Kimi K3, a model from Moonshot that the company says performs at a level comparable to leading American models.

Israel's middle path

Israel has largely followed the US approach, encouraging fast growth and avoiding heavy regulation, while treating AI as a driver of both economic growth and military strength. The government favors public funding for infrastructure, light regulation, and broad private-sector freedom as long as national security is not immediately at risk. Israel has decided to invest heavily in domestic supercomputing infrastructure to reduce dependence on US-based servers. The Innovation Ministry, the Israel Innovation Authority, and MAFAT, the Defense Ministry's Directorate of Defense Research and Development, are promoting a language model tailored to Israel for secure use in government ministries and the public sector.

Why this matters for government professionals

The three approaches show that AI policy is diverging sharply by country, and the rules governing access to advanced models can change quickly. Washington's shift from deregulation to targeted controls demonstrates that national security concerns can override commercial momentum overnight. Government professionals should expect more cases where access to advanced AI is restricted or conditioned on government approval, and should know the specific rules that apply to their own institutions. The gap between what AI can do and what governments will allow it to do is likely to widen, and public-sector employees will be on the front line of enforcing those boundaries.