The extraordinary move by the US to bar foreign access to Anthropic PBC’s best AI models underscores the Trump administration’s newfound willingness to exert control over a pivotal industry. It also reminds Silicon Valley that it’s working with an imperfectly understood technology with uncertain impact.
Washington has taken the unprecedented step of ordering the AI startup to disable access to its most advanced AI platforms for all foreign nationals. The US government issued the order after discovering it’s possible to “jailbreak,” or bypass the guardrails, of the Fable 5 AI model Anthropic released just days prior. The swiftness of the move triggered widespread concern over what the startup called a “disproportionate” response — and a warning that such an approach could “halt all new model deployments for all frontier model providers.”
Top Anthropic personnel are now in talks with officials in Washington about the specific security concerns that’ve been raised, people familiar with the matter said, asking to remain unidentified to discuss private deliberations. One of the key officials in negotiations is Treasury Secretary Scott Bessent, who’s sounded warnings for Wall Street about the potential dangers of frontier models such as Anthropic’s Mythos: a platform that theoretically discovers and exploits flaws in software quicker than humans.
The US government’s response marked the most significant incursion by officials into an AI company’s operations ever undertaken. It’s unfolding just as the largest AI startups, including Anthropic — already valued at more than $900 billion — seek to become publicly traded companies. The surprise controls threaten to set a precedent for major AI model developers, including OpenAI, Alphabet Inc.’s Google and Meta Platforms Inc.
The government is now willing to use extraordinary powers to compel AI developers to do its bidding on dealing with potential security threats — despite a June executive order that declared it wouldn’t force compliance on Silicon Valley. That earlier order declared the US wasn’t creating a licensing regime for AI models, but the Trump administration has moved to do exactly that.
“US frontier models are increasingly treated as strategic assets, with access tightly controlled and shaped by national security considerations,” said Gary Tan, a portfolio manager at Allspring Global Investments. That’s “a dynamic that is likely to persist as China continues to lag the US in compute.”
The debate is taking place against the backdrop of a race with China to dominate AI technology, as players like Alibaba Group Holding Ltd. and DeepSeek close the gap with their US rivals in performance and efficiency.
Anthropic released Fable 5 last week as the first public-facing version of its Mythos-class model. The company has repeatedly warned about the cyber capabilities of Mythos, and White House officials have taken those concerns seriously, creating a pathway through the voluntary executive order for US agencies to use Mythos and other cyber-capable models to patch weak spots in their own systems.
Washington’s fears may extend beyond just the ability to jailbreak and influence any one platform. Broadly, the worry also is that adversaries may try to steal the AI model weights — a valuable file of numbers that holds the key to the model’s capabilities. The move to bar Fable and Mythos coincides with growing concern in Washington that AI companies must guard their trade secrets against possible insider threats.
After decades of disseminating cutting-edge US technology around the world as a diplomatic and economic tool, Washington is now moving in the opposite direction. As with curbing the export of chips to geopolitical rivals such as China, the US is increasingly trying to preserve technology for home use, as a matter of national security.
“This shows the US racing to use export controls to rein in frontier Al risks,” said Stefanie Kam, an assistant professor at the China Programme, Institute of Defence and Strategic Studies at Singapore’s Nanyang Technological University.
“We can expect Washington to tighten curbs and treat Al exports as strategic leverage against Beijing. If the directive is narrowly targeted, US firms can adapt,” she said. “If it’s sweeping, it risks pushing innovation offshore while China advances.”
Last week, Senator Jim Banks, a Republican from Indiana who serves on a committee with export control oversight, called on Trump Administration officials to consider how the US could protect AI model weights from being stolen by China or other adversaries.
Policy objectives aside, Silicon Valley is likely to resist Washington’s attempts to exert control. From Meta to OpenAI, the biggest developers are racing to stake their claim on a technology that’s likely to reshape whole economies and industries.
Yet it’s impossible to test every permutation of an AI model’s potential capabilities, which leads to vague fears about threats and continuous scrutiny as the models keep improving.
“There may be important capability thresholds — such as AI systems that can automate large amounts of AI research and development in short periods of time — that pose new kinds of risks,” Banks wrote in a letter this month to officials including Bessent, National Cyber Director Sean Cairncross and Defense Secretary Pete Hegseth.
“Industry experts have likewise raised serious concerns about the difficulty of keeping such systems under human control.”
Photo: The Anthropic logo. Photographer: Gabby Jones/Bloomberg
Copyright 2026 Bloomberg.
Topics
USA
The most important insurance news,in your inbox every business day.
Get the insurance industry’s trusted newsletter
