Concerns raised by CEO Andy Jassy reportedly prompted the Trump administration to block foreign access to two advanced models, exposing tensions between national security, corporate partnerships and the global development of artificial intelligence.
WASHINGTON — A warning delivered by Amazon chief executive Andy Jassy to senior U.S. officials helped trigger an extraordinary government crackdown on Anthropic’s most advanced artificial-intelligence systems, abruptly cutting off access for foreign users and forcing the company to take the models offline worldwide.
The intervention followed tests conducted by Amazon researchers on Anthropic’s recently released Fable 5 model. According to people familiar with the discussions, the researchers were able to prompt the system into producing information that could potentially assist cyberattacks, despite safeguards intended to prevent such use.
Jassy subsequently raised the findings with senior members of President Donald Trump’s administration, including Treasury Secretary Scott Bessent. White House officials convened discussions, government security specialists examined the claims and Anthropic was instructed to correct the vulnerabilities or withdraw the affected technology.
The administration then went further.
Citing national-security concerns, the Commerce Department imposed an export-control restriction barring all foreign nationals from accessing Fable 5 and the more capable Mythos 5 model. The measure applied not only to users in countries regarded as strategic adversaries, but also to citizens of allied states and foreign employees working inside the United States.
Anthropic responded by disabling the two models globally, saying it could not immediately operate them while reliably separating eligible American users from everybody else.
The result was one of the most sweeping government interventions yet imposed on a commercially available AI system.
From Corporate Warning to Government Order
Amazon’s role in the episode is complicated by its exceptionally close commercial relationship with Anthropic.
The technology group is one of Anthropic’s largest investors and provides much of the cloud computing and specialised hardware used to train and operate its models. The companies recently expanded their partnership, with Anthropic committing heavily to Amazon Web Services and Amazon’s Trainium processors.
That relationship makes Amazon both a beneficiary of Anthropic’s growth and an influential gatekeeper in its technological development.
According to reports, Amazon’s internal researchers discovered that Fable 5 could be manipulated—or “jailbroken”—into identifying software weaknesses and supplying information that its safety rules were designed to withhold.
Jailbreaking generally involves structuring prompts in ways that persuade an AI system to ignore or circumvent its behavioural restrictions. Researchers routinely test advanced models for such weaknesses because they can allow users to obtain instructions involving malware, cyber intrusion, weapons or other dangerous activities.
Government officials treated Amazon’s discovery as evidence that foreign access to the models could pose an immediate security risk.
Anthropic disputed the severity of that assessment. The company said the government had not provided detailed technical evidence and argued that the identified weakness was narrow, involved already known vulnerabilities and did not give the model capabilities substantially beyond those available elsewhere.
The disagreement reflects a broader problem in AI regulation: officials may have to act before the scale of a threat is fully understood, but precautionary decisions can also produce enormous commercial and scientific disruption.
A Ban Extending to America’s Allies
The breadth of the order surprised even some specialists who support tighter controls on advanced AI.
Rather than targeting designated adversaries or specific overseas organisations, the restriction applied to every foreign national. That potentially included Canadian, British and European researchers employed by Anthropic, as well as trusted customers operating in allied countries.
The measure therefore treated access to the models in a manner resembling sensitive military technology.
Until now, Washington’s most prominent AI export controls have focused largely on advanced semiconductors, chipmaking equipment and computing infrastructure. Restricting the use of a specific model represents a significant expansion of that strategy.
The distinction matters because AI models are not physical products that can simply be stopped at a border. They are accessed through cloud platforms, application-programming interfaces and research networks involving international teams.
Enforcing nationality-based access may require identity checks, employee restrictions, technical segregation and changes to how cloud services are designed.
For Anthropic, disabling the models globally was the fastest way to comply. It was also commercially damaging.
Customers that had begun experimenting with Fable 5 lost access almost immediately. Research projects were disrupted, international developers were excluded and uncertainty spread across the company’s partnerships.
The restriction could also complicate Anthropic’s reported preparations for a public stock offering by creating new questions about regulatory risk and the company’s ability to serve a global market.
Security Concern or Political Escalation?
The order cannot be separated from Anthropic’s troubled relationship with the Trump administration.
Earlier disputes centred on the company’s limits governing military applications of its technology. Anthropic resisted allowing its models to be used for certain forms of autonomous weaponry and mass surveillance, positions that drew criticism from administration and defence officials.
The Pentagon subsequently moved against the company over alleged supply-chain concerns, creating tension with contractors and investors.
Against that background, Anthropic and some outside observers have questioned whether the latest decision was based solely on technical evidence or was influenced by wider political distrust of the company.
There is no public evidence that Amazon sought the broad foreign-access ban eventually imposed by the government. Jassy reportedly shared security findings, while the administration determined the scope of its response.
Nevertheless, the episode demonstrates how information exchanged privately between corporate executives and government officials can rapidly reshape an entire technology market.
It also highlights the influence held by large cloud providers. Amazon, Microsoft and Google do not merely host AI companies; they invest in them, supply their computing power, distribute their models and maintain their own competing AI interests.
A warning from one of these companies can therefore carry far greater weight than an ordinary external security report.
Amazon’s Conflicting Interests
Amazon has strong reasons to protect the security of models offered through its infrastructure.
A powerful system that could easily be exploited for cyberattacks might expose AWS customers, damage confidence in cloud AI and invite even more intrusive regulation.
At the same time, Amazon’s relationship with Anthropic creates an unavoidable appearance of conflicting interests.
Amazon has invested billions in the company and wants its technology to drive demand for AWS. Yet it is also developing its own AI systems, chips and services in a highly competitive market.
The shutdown of Anthropic’s most capable models could weaken an important partner while altering competition among leading AI laboratories.
There is no indication that Jassy raised the issue to gain a competitive advantage. The security concern itself was reportedly based on testing conducted by Amazon researchers. But the episode illustrates why the concentration of AI infrastructure among a few powerful companies creates governance challenges.
The same organisation can simultaneously be investor, supplier, distributor, security evaluator and competitor.
A New Era of AI Export Controls
The decision may establish a precedent extending far beyond Anthropic.
If advanced models are treated as controlled strategic assets, other American AI companies could face similar restrictions whenever their systems cross an undefined threshold of capability or risk.
That prospect could transform how frontier models are developed and released.
Companies may delay public launches, restrict international access from the outset or create separate versions for American and overseas users. Foreign researchers could be excluded from core development teams, while multinational businesses may become hesitant to build products around systems that Washington can suddenly withdraw.
The policy could also accelerate efforts by Europe, China and other regions to develop independent AI infrastructure.
Governments and companies may conclude that relying on American models leaves them vulnerable to unilateral restrictions imposed with little notice. A policy intended to protect U.S. technological leadership could therefore encourage other countries to reduce their dependence on American providers.
Supporters of the order will argue that such costs are justified if frontier AI can materially increase the speed and sophistication of cyberattacks.
AI systems are increasingly capable of analysing code, identifying security flaws and automating technical tasks. Even where a model does not independently create a novel weapon, it may lower the expertise needed to exploit existing vulnerabilities.
Critics counter that restricting one company’s models does little if comparable capabilities remain accessible through competitors or open-source systems.
A narrowly focused ban may therefore disrupt legitimate users without meaningfully preventing malicious actors from obtaining similar assistance elsewhere.
The Challenge of Regulating at Machine Speed
The Anthropic episode shows how quickly the relationship between AI companies and governments is changing.
A private security test was communicated to senior officials. Emergency discussions followed. Within days, a federal order altered access to technology used by researchers and businesses around the world.
Such speed may become increasingly common as models grow more capable.
Traditional regulation moves through public consultation, legislation and judicial review. Frontier AI develops on a much faster cycle, creating pressure for governments to intervene through executive orders, export controls and emergency national-security powers.
That approach allows officials to respond rapidly, but it offers companies and users limited transparency or opportunity to challenge the evidence.
The central question is no longer whether advanced AI will be regulated. It is who will define the threshold at which a model becomes too dangerous for international use—and what evidence will be required before access is withdrawn.
For Anthropic, the immediate priority is to address the government’s concerns and restore access without compromising its safety principles.
For Amazon, the incident underscores the responsibility that comes with its position at the centre of the AI industry.
And for the Trump administration, the order marks a decisive expansion of American technology controls—from restricting the chips that power artificial intelligence to deciding who may use the intelligence itself.
