The United States has stepped back from a hard restriction, but not from the logic of control itself. The lifting of export curbs on Anthropic’s newest models, Fable 5 and Mythos 5, is not a victory of industry over the state. It is a signal of a new regime for frontier artificial intelligence.
Less than three weeks earlier, access to these models had effectively been halted over national security risks. The reason was direct: Washington did not want powerful AI systems to quickly end up in the hands of foreign users linked to countries it considers strategic threats.
Anthropic disabled both models for all users because it had no reliable way to verify nationality in real time. That episode exposed a weak point across the industry: global AI services can be launched instantly, while state rules still require borders, identification and accountability.
According to Daycom’s earlier analysis, the Anthropic case matters not only for one developer. It shows how U.S. authorities are moving from abstract discussions about AI safety to concrete mechanisms for access, suspension, testing and possible reimposition of restrictions.
Mythos 5 is especially sensitive because it was designed to detect cybersecurity vulnerabilities. In a normal corporate setting, such a model can help identify weaknesses in systems faster. In the hands of a malicious actor or hostile intelligence service, the same capability can become an attack tool.
Fable 5 is aimed at a broader public and has stronger safeguards. Yet one of the central problems emerged around that model. Amazon researchers found a way to bypass its restrictions, allowing it to identify software vulnerabilities and, in one case, generate code demonstrating how one could be exploited.
Anthropic has since introduced a new safeguard to block the behavior described in that report. Requests that enter a dangerous zone are now redirected to Opus 4.8. That may frustrate some users, but the inconvenience has become the price of making the model’s other capabilities widely available.
The most important part of this story is not a technical detail, but an admission of limits. Even companies that build their reputation on safety no longer promise absolute resilience. Making an AI model fully robust against jailbreaks — attempts to bypass safeguards — is nearly impossible.
That is why the market is entering a new phase. Developers once competed mainly on power, speed, context length, coding quality and reasoning ability. Now a competitive advantage also means proving to governments, customers and partners that a model does not open uncontrolled access to dangerous actions.
Anthropic is expanding cooperation with the U.S. government and giving designated government partners early access to its models. That changes the nature of frontier AI launches. A public release is no longer only a company decision involving marketing and server capacity. It is becoming a matter of security clearance.
Another track involves shared standards with Amazon, Microsoft, Google and other partners in the Glasswing program. The goal is to build systems for assessing jailbreaks, fixing them and ranking their severity. The industry is trying to create a common language for a risk that was often described chaotically.
That is critical for the market. Without shared standards, every company can call its model safe, while users, governments and businesses lack a common basis for comparison. For frontier AI, that uncertainty becomes both a political and commercial risk.
U.S. Commerce Secretary Howard Lutnick left the department the right to revisit the decision and reimpose licensing requirements if circumstances change or Anthropic fails to meet its commitments. That means the approval is neither unconditional nor final.
In effect, Washington is creating a model of conditional trust. A company receives access to the market, but accepts obligations to report malicious activity, work on safety protocols and maintain contact with the state. If that balance breaks, restrictions can return quickly.
This is the new reality for American AI companies. They must remain global businesses, innovation laboratories and elements of national security at the same time. The boundary between a commercial product and a strategic technology is now fully blurred.
For Anthropic, the story is especially sensitive because of its tense relationship with Donald Trump’s administration. Earlier, the Pentagon identified the company as a supply-chain risk after it refused to allow its models to be used for mass domestic surveillance or fully autonomous weapons systems.
That conflict reveals a broader dilemma: the government wants access to the best models, but not every developer is willing to place them under any state scenario. Companies want to preserve ethical boundaries while also needing regulatory approval, contracts and political protection.
Similar restrictions have also affected OpenAI, which delayed the full public launch of GPT-5.6 at the U.S. government’s request and limited access to a small group of vetted partners. This is no longer an isolated incident. It is the formation of a new procedure for frontier models.
Trump’s executive order created a voluntary framework under which developers can offer the government access to advanced models before releasing them to trusted partners. Formally, it is voluntary. In practice, refusal may carry a high political cost for companies operating at major scale.
The central question now extends beyond the United States. If American models face increasingly strict oversight, how should comparable systems from China or other countries with weaker safeguards be handled? Domestic regulation does not solve global asymmetry.
Washington can restrain its own companies, but it cannot automatically impose the same rules on the world. If excessive control slows American developers while competitors move more aggressively, security policy may come into conflict with technological advantage.
That is why the decision on Anthropic looks like a compromise. The government is not keeping the models blocked indefinitely, but it is not returning the market to the old freedom of launches. Access is allowed only after safeguards, cooperation, standards and a direct possibility of renewed intervention.
For users, this means a less predictable but safer landscape. New models may arrive with pauses, restrictions, vetted access groups and redirection of dangerous requests. From a business perspective, that is inconvenient. From a state perspective, it is becoming unavoidable.
Frontier artificial intelligence can no longer be launched like an ordinary digital product. Its ability to write code, identify vulnerabilities, automate analysis and assist with complex operations has made it part of the security infrastructure. That is what the U.S. government has now recognized.
The lifting of restrictions on Fable 5 and Mythos 5 therefore does not end the dispute. It opens a new phase in which every major AI release will pass through the same questions: who gets access, what safeguards are in place, who tests the model and how quickly the state can stop its spread.
In this new formula, the future of AI will not be shaped by engineers alone. It will also be shaped by lawyers, intelligence agencies, ministries, cloud platforms, safety researchers and users themselves, whose attempts to break the system have already become part of the testing process. The era of uncontrolled releases is ending.
