The U.S. government has lifted export controls on Anthropic’s Fable, bringing the model back after a two-week policy whiplash that left both Fable and Anthropic’s Mythos disabled for all users. The reversal wasn’t instantaneous. Mythos was restored first on Friday evening, then Fable late on Tuesday, after controls imposed two weeks ago forced Anthropic to shut off access.

For anyone trying to build products, defend systems, or forecast risk, this matters fast. Fable can carry out lengthy tasks autonomously, so when it goes dark, workflows do too. And for cyber defenders, Mythos was especially valuable because researchers have been eager to use it to find security flaws and patch them before attackers get access to models of equal capability. The good news is that Anthropic’s model availability is back. The bad news is what this episode reveals about how shaky the “rules of the road” still are for frontier AI in the United States.

Under the surface, the U.S. is operating what is essentially a licensing regime for frontier AI models while officially denying that it is. That licensing behavior has been, by the story’s account, largely ad hoc, with opaque rules apparently invented on the fly by various U.S. government officials. The result is a familiar executive nightmare: policy that can change mid-flight, even if you technically did everything right on launch day. When governments behave like this, companies do not just adjust their tech plans. They redesign their risk models, procurement strategies, and “what is our fallback if the model suddenly disappears” playbooks.

Recent reporting suggests the chaos may be starting to get a structure. According to a story in the Financial Times, the U.S. is working with leading AI labs on an explicit set of “voluntary standards” that frontier AI labs can meet, at least regarding cybersecurity, to have a reasonable expectation the government won’t object to a model’s public release. In parallel, Anthropic announced it is working with the U.S. government on a shared framework for assessing how risky a jailbreak to a model’s guardrails poses. Anthropic said it worked with Amazon, Microsoft, Google, and what it described as “other Glasswing partners” initially, and welcomed others to join. Glasswing is the name Anthropic gave to the coalition of critical infrastructure companies allowed access to Mythos.

The governance angle is not just bureaucratic. It’s about incentives and trust between rivals. The story notes it was “interesting” that Anthropic did not include its rival, OpenAI, in the initial group from the get-go. That tiny detail hints at a bigger reality in frontier AI: collaboration on standards and shared risk assessment is easier when everyone is aligned. It is harder when competitive dynamics and regulatory access create unspoken power struggles.

Still, damage has already happened. The export controls were temporary, but the strategic lesson sticks: it can be “strategically unwise” to count on American frontier models for anything essential. That perspective is especially common in Europe, where Bea has reported similar concerns. Even in the U.S., the story points to more enterprises talking about open source models as fallback options.

Which open source models, though, is where the plot thickens. The world’s most capable open models currently come from Chinese AI companies. Western businesses can often download those models and run them on their own cloud infrastructure to reduce data leakage risk back to China. But reputational risk is still real. There is also the risk that U.S. policy could shift in a way that blocks American firms from using Chinese models, a scenario some have suggested is a likely policy outcome. In other words, “fallback to open source” may not be a single switch you flip. It may be a multi-variable decision involving law, reputation, and future access.

There’s also the cyber reality check. 2 had, according to one cybersecurity research firm, equalled the capabilities of Anthropic’s Mythos. But the story cautions that it likely did not equal Mythos across all capabilities as described to the Journal. The research firm’s framing suggests GLM-5.2 could spot many of the same software vulnerabilities, but Mythos is special for autonomously chaining vulnerabilities into working exploits and using those exploits to conduct hacks. The story says there is no indication GLM-5.2 can do that, but adds a timeline risk: it’s probably only months before some open source model can.

Guardrails get weaker when models become more accessible. The piece explains that AI misuse prevention today often relies on guardrails and classifiers, sometimes smaller models trained to screen prompts. With open source models, those classifiers can be stripped away, and guardrails can be jailbroken. The story also states that if an attacker has access to model weights, which open source can provide, there is always a jailbreak that can overcome trained-in guardrails. This is part of why the Five Eyes intelligence agencies recently warned of an imminent cyber threat from advanced AI models.

Meanwhile, OpenAI’s Sam Altman is renewing calls for a U.S.-led international AI governance regime with shared standards across countries, in exchange for shared access to technology. The story notes it’s unclear whether such an effort would include China. One idea discussed is basing the initial governance regime out of the G7, which does not include China. Still, momentum is building toward more transparent AI regulation domestically and internationally. The open question is whether it arrives in time.

If you’re an executive, investor, or operator, the takeaway is blunt: the frontier AI market is moving toward frameworks, but today’s environment rewards contingency planning. Export controls may loosen, models may return, standards may emerge. But the underlying uncertainty about how quickly policy can pivot is already forcing enterprises into fallback strategies, broader model sourcing, and tighter governance of cybersecurity risk.