Anthropic’s Fable model is back. But U.S. AI policy is still a mess

The U.S. government has lifted export controls on Fable and Mythos. But there's still no clarity about the rules that will govern future AI model releases.

Hello and welcome to Eye on AI. It’s Jeremy here, filling in for Bea, who usually writes the Thursday newsletter. In this edition:

• Anthropic’s Fable is back. But U.S. AI policy is still a mess.
• OpenAI wants the U.S. government to take a 5% stake in the company
• And OpenAI reportedly scores a breakthrough in compute efficiency
• Plus Meta stock soars on plans to launch a cloud computing business

The biggest AI news of the past week has been the government’s decision to roll back the export controls it had imposed two weeks ago on Anthropic’s Mythos and Fable models. Those controls had resulted in Anthropic having to disable both models for all users. The government first reversed course on Mythos on Friday evening and then on Fable late on Tuesday. You can read more about the Fable decision here from Fortune’s Tristan Bove.

The decision will be a relief for Anthropic and its investors, and for many people who were hoping to use Fable, which can carry out lengthy tasks autonomously. (Whether Anthropic investors should really be happy is another matter; there’s an argument to be made that Anthropic might have avoided this two week crisis with a different political comms strategy, an idea that I explore in this Fortune story.) It will also cheer cyber defenders who have been eager to use Mythos to find security flaws and patch them before attackers have access to models of equal capability.

But the latest decision to lift the export controls still leaves American AI policy in something of a mess. The U.S. is continuing to operate what is essentially a licensing regime for frontier AI models, while officially denying that this is the case. This licensing regime is also almost completely ad hoc, with opaque rules apparently being invented on the fly by various U.S. government officials.

Now, there are reports that this may be about to change. 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 when it comes to cybersecurity—to have a reasonable expectation that the U.S. government won’t object to a model’s public release. In addition, Anthropic announced that it is working with the U.S. government on a shared framework for assessing the level of risk that a jailbreak to a model’s guardrails poses. Anthropic said it was working with Amazon, Microsoft, Google, and what it described as “other Glasswing partners” initially on this framework but welcomed the participation of others. (Glasswing is the name Anthropic has given to the coalition of critical infrastructure companies that have been allowed access to Mythos.) It was interesting—and perhaps shows the level of distrust between Anthropic and OpenAI—that they did not include their rival in that group from the get-go.

The damage is done

Still, damage from the export control decision has already been done and cannot be undone. The decision to impose the export controls, even temporarily, has forced potential customers of American frontier AI models to recognize that it might be strategically unwise to count on these models for anything essential. At the very least, fall-back options are needed. This view is especially prevalent in Europe, as Bea has reported. But even in the U.S., a lot more enterprises are now talking about open source models.

The question is, which open source models? The world’s most capable are all from Chinese AI companies, which presents Western businesses with a dilemma. While a company can download these models and run them on their own cloud infrastructure to eliminate any risk of data leakage back to China, using a Chinese model still presents reputational risks—and the risk that the U.S. government might seek to cut off American firms from using Chinese models, as some have suggested is a likely policy outcome.

As open source models become more capable, governments are going to face a real dilemma about what to do about them. The Wall Street Journal reported earlier this week that Zhipu AI’s GLM-5.2 model had, according to one cybersecurity research firm, equalled the capabilities of Anthropic’s Mythos. Only, from what the researchers told the Journal, it didn’t sound like GLM-5.2 actually equalled Mythos in all its capabilities. Rather, it seemed GLM-5.2 was able to spot many of the same software vulnerabilities as Mythos. But that is true of several other publicly-available models too. What makes Mythos special is its ability to autonomously work to chain vulnerabilities together into working exploits and use those exploits to carry out hacks. There’s no indication that GLM-5.2 can do that. But it is probably only a matter of months before some open source model can.

No way to guardrail open source

Worse, the ability to prevent AI models from being used for cyber attacks today largely depends on guardrails and using classifiers—or other, small AI models that try to screen the prompts being fed to the larger model and disallow ones that seem suspect. But with open source models, these classifiers can easily be stripped away and the guardrails that might be trained into a model can also easily be jailbroken. In fact, researchers have shown that if an attacker has access to a model’s weights, which is the case for open source, then there is always a jailbreak that can be found that will overcome any trained-in guardrails.

So we are almost certainly heading for trouble. This is why the Five Eyes intelligence agencies recently warned of an imminent cyber threat from advanced AI models. It is also why OpenAI’s Sam Altman is renewing calls for a U.S.-led international AI governance regime that would see at least Western governments cooperate on shared standards around AI, in exchange for getting shared access to the technology. While it’s 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), it might still provide a framework for safe sharing of powerful models to help Western countries defend against AI-powered attacks.

So there is momentum towards transparent AI regulation both domestically and internationally. But whether that regulation will arrive in time is another question.

With that, here’s more AI news.

Jeremy Kahn
jeremy.kahn@fortune.com
@jeremyakahn

This story was originally featured on Fortune.com