The U.S. government has lifted export controls on Anthropic's Fable and Mythos AI models after a two-week shutdown, but the crisis has exposed a fundamental problem: American AI policy operates as an ad hoc licensing regime with opaque rules that shift without warning. The rollback, which came after the government reversed course on Mythos on Friday and Fable late Tuesday, offers temporary relief to Anthropic and its users. However, the damage to trust and business planning may be irreversible.

The export control decision forced potential customers of American frontier AI models to confront an uncomfortable reality: relying on these tools for anything essential could be strategically dangerous. Companies now recognize they need fallback options, a concern especially prevalent in Europe but increasingly common in the U.S. as well. This shift is driving enterprises to explore open-source models, which introduces a new set of geopolitical tensions.

Why Is U.S. AI Policy Creating a Crisis of Confidence?

The core problem is that the U.S. continues to operate what amounts to a licensing regime for frontier AI models while officially denying this is the case. Government officials appear to be inventing rules on the fly, with no transparent standards or predictable process. This opacity has real consequences: companies cannot plan long-term strategies when the rules change without notice.

There are signs this may be changing. According to reporting, the U.S. is working with leading AI labs on an explicit set of "voluntary standards" focused on cybersecurity, which frontier AI labs can meet to have reasonable expectations that the government won't object to a model's public release. Additionally, Anthropic announced it is working with the U.S. government on a shared framework for assessing the risk that a jailbreak to a model's guardrails poses.

Anthropic stated it was collaborating with Amazon, Microsoft, Google, and other members of what it calls the "Glasswing" coalition of critical infrastructure companies on this framework. Notably, OpenAI was not included in the initial group, a decision that reflects the level of distrust between the two companies.

How Are Companies Responding to Policy Uncertainty?

• Exploring Open-Source Alternatives: Enterprises are increasingly discussing open-source models as fallback options, recognizing that American frontier models may become unavailable without warning.
• Evaluating Chinese Models: The world's most capable open-source models come from Chinese AI companies, presenting Western businesses with a dilemma between data security and reputational risk.
• Building Redundancy: Companies are developing strategies to run models on their own cloud infrastructure to eliminate data leakage risks while maintaining operational continuity.

The shift toward open-source models raises a critical question: which ones? Chinese AI companies have released some of the world's most capable models. While companies can download these models and run them locally to prevent data from flowing back to China, using a Chinese model still carries reputational risks and the possibility that the U.S. government might restrict American firms from using such tools.

Recent reporting noted that Zhipu AI's GLM-5.2 model had, according to one cybersecurity research firm, equaled the capabilities of Anthropic's Mythos. However, the comparison may be overstated. GLM-5.2 appears able to spot many of the same software vulnerabilities as Mythos, but what makes Mythos special is its ability to autonomously chain vulnerabilities together into working exploits and carry out hacks. There is no indication that GLM-5.2 can do that, though it is probably only a matter of months before some open-source model can.

What Makes This a Cybersecurity Crisis?

The ability to prevent AI models from being used for cyberattacks today largely depends on guardrails and classifiers, which are small AI models that screen prompts and disallow suspicious ones. With open-source models, these classifiers can easily be stripped away and guardrails can be jailbroken. Researchers have shown that if an attacker has access to a model's weights, which is the case for open-source models, there is always a jailbreak that can be found to overcome any trained-in guardrails.

This is why the Five Eyes intelligence agencies recently warned of an imminent cyber threat from advanced AI models. The problem is structural: as open-source models become more capable, governments face a real dilemma about what to do. They cannot easily prevent the spread of powerful models once they are released publicly, yet they also cannot ignore the security risks these models pose.

OpenAI CEO Sam Altman is renewing calls for a U.S.-led international AI governance regime that would see Western governments cooperate on shared standards around AI in exchange for getting shared access to the technology. While it is unclear if Altman's proposal would include China, one of his ideas is to base the initial governance regime out of the G7, which does not include China. Such a framework might provide a way for Western countries to safely share powerful models to help defend against AI-powered attacks.

There is momentum toward transparent AI regulation both domestically and internationally. However, whether that regulation will arrive in time to prevent the security risks posed by open-source models remains an open question. The two-week crisis over Fable and Mythos has made clear that the current ad hoc approach is unsustainable and that a more coherent policy framework is urgently needed.

" }