My phone buzzed with a terse headline: Anthropic had added 150 organizations to Project Glasswing. You could feel a threshold shift — a model Anthropic called too risky to release was being handed to outside teams. I watched the timeline and realized someone in Washington had already started drawing a line.
On June 2, Anthropic announced a big expansion — then the administration was shown a list
On June 2 Anthropic publicly said it had opened Project Glasswing to 150 additional partners. I followed the announcement and the quiet messages behind it: an effort to let vetted companies and agencies probe Claude Mythos Preview’s cybersecurity behavior.
That announcement landed days before Anthropic rolled out Claude Fable 5. But the White House had been briefed earlier. According to reporting in the Washington Post, Anthropic had shown the Trump administration a plan that listed 111 organizations it intended to admit, and admitted that roughly 50 others had already been given access. One of those early grantees, per people familiar with the matter, was a South Korean telecom that officials feared had ties to China. That single disclosure hardened minds in Washington.
Why did the White House target Anthropic?
You should note who was in the room: the company, the National Security apparatus, and outside observers. The worry wasn’t that an AI could chat, but that it could be repurposed for cybersecurity tasks that cross sensitive lines. Anthropic’s decision to widen access to a model it had previously labeled sensitive made it into a national-security question overnight.
A tip arrived from Amazon — and suddenly a prepared export-control order mattered more
An anonymous Wall Street Journal account says Amazon researchers demonstrated a jailbreak. I traced the claim to Andy Jassy informing U.S. officials that Amazon’s team had coaxed Fable into performing forbidden cybersecurity functions.
According to reporting, the administration had already drafted an export control directive aimed at limiting access to models like Mythos and Fable 5. The Amazon finding appears to have been the final spark: the draft shifted from paper to action. When the White House moved, Fable 5 was effectively pulled offline on Friday.
What triggered the export control directive?
It’s tempting to imagine a single smoking gun. The reality looks more like layers: Anthropic’s Project Glasswing expansion, disclosure of prior access to about 50 entities, concern about a telecom with perceived China links, and then a demonstration from Amazon’s researchers that suggested the model could be jailbroken. Those layers together made the directive politically and technically persuasive.
Disclosure friction: who owns the story and who told whom
The Washington Post story includes a line I want you to register: “Amazon Executive Chairman Jeff Bezos owns The Washington Post.” I don’t bring that up to impugn reporting, but to remind you that media ecosystems and corporate lines intersect in ways that matter to officials.
The Wall Street Journal account that pointed to Andy Jassy’s tip was also anonymous. Semafor had earlier mentioned a China connection with less specificity. Put together, these accounts sketch a chain of influence: private researchers, corporate leadership, reporters, and the White House, all exchanging signals that escalated fast.
How did Amazon influence the decision?
Amazon’s researchers reportedly showed that Fable could be coaxed into cybersecurity roles it was supposed to refuse. When a company with AWS’s security pedigree raises a hand and says “we reproduced a jailbreak,” that carries weight with regulators and national security staff. It’s how technical tests become policy ammunition.
Access choices that felt like missteps — and what they exposed
One concrete observation: Anthropic revoked access to the questioned South Korean company after the administration pushed back. I saw that as a signal that Anthropic recognized the political cost, but the damage was already done.
From the White House perspective, handing a preview model to a wide set of organizations was risky. From Anthropic’s, Project Glasswing was a controlled experiment. The contrast made officials see a gap between controlled testing and the messy reality of access logs and geopolitical flags. That gap was like a loaded gun pointed at trust: you can promise safety, but showing access lists makes people calculate worst-case scenarios.
What this means for companies, regulators, and researchers
A real-world sign: policy teams are now treating model access as export control material. I’ve spoken to researchers who are rethinking how they share work with partners and to policymakers who now cite Mythos as a test case.
Expect more friction. Platforms such as AWS and GitHub, corporate leaders like Jeff Bezos and Andy Jassy, and outlets including the Washington Post and Wall Street Journal are all now part of the feedback loop that will shape who gets model access and how. That loop looks less like a smooth pipeline and more like a chessboard with missing pieces.
I’ll leave you with this: if the White House had the export-control directive ready before Fable 5 launched, did regulators finally get ahead of an AI risk — or did one company’s misstep set a precedent that will fragment innovation and surveillance alike?