Anthropic’s Fable and Mythos models are back. But U.S. AI policy is still a mess.

• 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
[email protected]
@jeremyakahn

FORTUNE ON AI

The 33-year-old executive Satya Nadella is trusting to fix Microsoft’s Copilot AI assistant—by Sebastian Herrera

How foodservice giant Sodexo is embracing AI and robotics to reshape the kitchen—by John Kell

Cisco is rolling out AI agents to every single one of its 90,000 employees—by Sheryl Estrada

OCBC rolls out its ‘avatar banking’ platform with ‘Wendy’ and ‘Wayne,’ two virtual financial advisors, as banks integrate AI into wealth management—by Angelica Ang

AI IN THE NEWS

OpenAI discusses U.S. government taking a 5% equity stake. OpenAI has reportedly discussed the possibility of giving the U.S. government a 5% equity stake in the fast-growing AI company ahead of its likely IPO, according to a story in the Financial Times. The story says that other leading AI companies would also be asked to make a similar equity contribution to a sovereign wealth fund, which might be modeled on the fund Alaska uses to ensure that state’s citizens share in the state’s oil wealth. OpenAI CEO Sam Altman has argued that public ownership could help share the economic benefits of AI, helping to mitigate the negative views of the technology that most American voters currently have. It is also no coincidence that such a deal could strengthen OpenAI’s relationship with the Trump administration and possibly with future U.S. governments. OpenAI may also be hoping to head off a somewhat similar proposal from Vermont Senator Bernie Sanders which would force AI companies to contribute as much as 50% of their equity to a sovereign wealth fund.

OpenAI reported compute efficiency breakthrough drives down chip stocks. OpenAI engineers have reportedly developed new inference optimizations that more than halved the cost of running AI models, dramatically reducing the number of Nvidia GPUs needed to serve some ChatGPT traffic. That is according to a story in tech publication the Information that cited one unnamed source it said was familiar with what OpenAI engineers had said about the breakthrough. Although the story contained no details on exactly how OpenAI had achieved this alleged breakthrough, the report contributed to a sharp sell-off in the shares of computer chip makers on Wednesday. Shares in some companies, such as memory chip maker Micron, fell more than 10%, while the Philadelphia Semiconductor Index overall lost more than 6%. The Information noted that while the efficiency gains could help OpenAI lower prices, potentially gaining marketshare from rivals Anthropic and Google, it might also help the company boost its profit margins ahead of a likely IPO. 

Anthropic debuts Claude Sonnet 5. The company described it as its most capable Sonnet-branded model to date, with significantly stronger performance in coding, agentic use cases, and professional knowledge work. The company said the new Sonnet’s performance approaches that of its flagship Opus models, but at a lower cost. The company says the model is designed to be more autonomous—able to plan, use tools such as browsers and terminals, and complete complex, multi-step tasks more reliably—and includes improved safety features, including greater resistance to prompt injection attacks and malicious requests. Claude Sonnet 5 is being rolled out across Anthropic’s Claude plans and APIs with introductory pricing that is nominally the same as for its Sonnet predecessor, Claude Sonnet 4.6. (However, at least one independent expert noted online that the new Sonnet generates considerably more output tokens for a given prompt than its predecessor, meaning that the new model may be 30% more expensive for many use cases.) You can read Anthropic’s blog on the new model here.

Anthropic also debuts Claude Science. The company has introduced Claude Science, a product that it described as “a beta AI workbench” that can handle scientific literature review, data analysis, coding, visualization, manuscript writing, and high-performance computing in a single environment, aiming to streamline the end-to-end workflows that many scientists use. The new product includes features for keeping results reproducible and auditable, Anthropic said. It also includes domain-specific agents and more than 60 preconfigured scientific tools and connectors for fields such as genomics, proteomics, structural biology, and cheminformatics, and can manage compute jobs across local machines, lab clusters, and cloud GPUs while allowing sensitive data to remain on researchers’ own infrastructure. Early users report substantial productivity gains on tasks ranging from literature reviews to genomics and drug discovery, according to the company. You can read their blog on the new product here.

Likely next U.K. Prime Minister considers big shift in AI policy. A report in the Financial Times says that advisors to Andy Burnham, the Labour politician who is likely to succeed Keir Starmer as British prime minister by September, are developing a new U.K. AI strategy that would try to move the country away from dependence on large U.S. technology companies. The strategy would instead aim for increasing British tech sovereignty, support for domestic businesses, worker retraining, and stronger oversight of digital markets. The proposed approach may call for greater accountability over U.K. data centers, a review of the AI Growth Zones that Starmer created under his AI strategy, and policies that better address AI’s impact on workers and local communities. It’s unclear, however, how easily the U.K. could establish itself as an AI leader without reliance on American tech.

EYE ON AI NUMBERS

9%

That’s how much Meta’s stock climbed on Wednesday after reports emerged that the social media giant planned to launch a cloud computing business that would sell excess AI computing capacity. The new business line could help the company to recoup the huge sums—an estimated $145 billion this year alone—it is spending to build out AI infrastructure. Meta itself declined to comment. The news also hammered the shares of other AI cloud providers, particularly neoclouds such as CoreWeave, because of fears about increased competition. Read more from Bloomberg here.

AI CALENDAR

July 6-11: International Conference on Machine Learning (ICML), Seoul, South Korea.

July 7-10: AI for Good Summit, Geneva, Switzerland.

Aug. 4-6: Ai4 2026, Las Vegas.

Nov. 16-17: Fortune 500 Innovation Forum, Detroit. Apply here to attend.

Dec. 6-12: Neural Information Processing Systems (Neurips) conference. Sydney, Australia.

Dec. 7-8: Fortune Brainstorm AI, San Francisco. Apply here to attend.