The conversations took place in Washington this week, and their consequences reached the artificial-intelligence industry within days. Andy Jassy, Amazon’s chief executive, told senior officials in the Trump administration that he had safety concerns about the most advanced models built by Anthropic, the AI startup whose systems run largely on Amazon’s cloud, according to people familiar with the matter. Shortly after those talks, the U.S. government directed Anthropic to disable Claude Fable 5 and Claude Mythos 5 around the world.
The Wall Street Journal and The Information reported the sequence, and it has already become a study in how cloud providers, investors and regulators now share the industry’s most consequential decisions. Amazon occupies both sides of the table. Its AWS unit is Anthropic’s principal cloud partner and the channel through which thousands of businesses buy access to Anthropic models. Jeff Bezos, Amazon’s founder, has backed the startup through personal investment vehicles, tying Amazon’s fortunes to Anthropic’s rise. Mr. Jassy’s conversations in Washington sit awkwardly beside that arrangement.
Anthropic executives learned of the order by phone on a weekday morning. It covers two systems. Mythos 5, a research model that can identify security vulnerabilities across major operating systems and browsers, had been restricted to about 50 organizations since its release. Fable 5, a guardrailed version built for commercial customers, had been public for only three days, and Amazon had just begun offering it to enterprises through Bedrock, its AI platform. The speed of the reversal stunned engineers inside the company, people familiar with the matter said, because Fable 5 had passed the internal safety reviews that Anthropic treats as its most sacred process.
Anthropic said in a statement that the government supplied only “oral evidence” of what the company described as a “narrow, non-universal jailbreak” against the models. It disputed the basis for the action, arguing that the underlying capability already exists in OpenAI’s GPT-5.5 and is used by cybersecurity firms to find and patch flaws before criminals can exploit them. The company said it is complying with the order while pressing its objections, and that it has shared its own technical assessments with the agencies involved.
The dispute is the first time Washington has ordered a developer to shut down specific AI models, and it lands at a delicate moment for Anthropic. The company built its reputation on caution, publishing detailed release policies and rationing access to its most capable systems, a posture that won it enterprise clients and government work. Fable 5 was the payoff of that strategy: a frontier model judged safe enough for general commercial release. Its removal, three days in, leaves customers who built on it scrambling and raises questions about how durable any AI product launch can be when regulators can reverse it on short notice.
Amazon’s role makes the episode harder to read. The company is simultaneously Anthropic’s biggest business partner, a backer of its growth and, through its chief executive, a voice in Washington that preceded a ban on its partner’s products. People familiar with the matter said Mr. Jassy’s concerns focused on the models’ potential for misuse and that he did not ask for a shutdown. Even so, executives across the industry noted the timing: the concerns surfaced as Amazon was preparing to sell Fable 5 at scale.
The two companies’ financial ties deepen the puzzle. Amazon has committed up to $8 billion to Anthropic since 2023, making it one of the startup’s largest shareholders, and Anthropic has in turn become one of AWS’s most important AI customers. A ban that removes a flagship product from the platform, even temporarily, costs both companies. Anthropic, valued at $965 billion in recent private-market rounds and widely expected to follow SpaceX into the public market, now carries a new line of risk in its file: a government order that can switch off a product overnight.
The episode also tests how Washington treats the people who raise alarms. Mr. Jassy’s concerns, as described by people familiar with the matter, were delivered privately and in the language of risk management, not accusation. Yet the outcome, a global shutdown, gives other executives reason to think twice before carrying similar messages. The line between responsible disclosure and handing regulators a weapon against a rival is now blurry, analysts said.
For the wider industry, the episode establishes a template. Washington has shown it can act directly against a specific model rather than through broad rules, and a cloud giant has shown it can shape the conversation before regulators act. Companies building frontier models now face a question they rarely had to ask: which part of the chain, from the laboratory to the data center to the White House, can decide whether a product lives or dies.
Analysts said the episode is likely to push AI developers toward more conservative release policies until the legal questions are settled. Anthropic had already positioned itself as the industry’s cautious lab. Caution, after this week, may become the default for everyone else.
