Project Glasswing Explained: Anthropic Ran Authorized Vulnerability Tests With U.S. Intelligence Before the Government Shut Down Its Models

Since Anthropic disabled Mythos 5 and Fable 5 for all customers earlier this month to comply with a Trump administration directive restricting foreign national access, the central question has been: what exactly happened, and why did the government move so fast?

The AP, cited by CNBC, now has a partial answer.

What Project Glasswing Actually Was

A U.S. official, speaking anonymously to the Associated Press, confirmed that the vulnerability findings came out of a structured program called Project Glasswing. The initiative brought together Anthropic and U.S. intelligence agencies specifically to test whether the Mythos model posed risks to critical software infrastructure, national security, and the economy.

The official was careful about scope. Mythos identified vulnerabilities in classified systems within hours. It did NOT exploit them within that time. Those are two materially different things, and conflating them would overstate what the test demonstrated.

That distinction matters, because the public record to this point has been driven largely by Senator Mark Warner's characterization at a June 11 Senate Banking Committee hearing. Warner said, "This tool broke into almost all of our classified systems, not in weeks but in hours," attributing that assessment to General Joshua Rudd, head of the NSA and U.S. Cyber Command. Warner's phrasing implied active penetration. The official's clarification to AP suggests the more precise description is: the model mapped where the doors were, not that it walked through them.

The NSA declined to comment when CNBC reached out. Anthropic also declined to comment.

The Administration's Response and Anthropic's Compliance

The Trump administration issued a directive requiring Anthropic to block foreign nationals from accessing Mythos 5 and Fable 5. The directive came roughly ten days after President Trump signed an executive order establishing a voluntary framework for AI developers to submit their most advanced models for national security vetting before public release, according to CNBC.

To comply, Anthropic pulled the models entirely for all global customers, not just foreign nationals. The company said it did not believe the government's concern justified the steps it was being asked to take, but complied anyway.

That gap—complying while publicly disagreeing—is the clearest signal yet that Anthropic views the directive as disproportionate. The company released Fable 5 broadly earlier this month. That model is described as a limited version of the more restricted Mythos.

The Cybersecurity Industry's Counterargument

The strongest case against the administration's directive comes from within the security industry itself.

More than 100 cybersecurity executives and researchers, including leaders from Adobe and Nvidia, sent a letter to the Trump administration asking it to lift the restrictions. Their argument: models capable of finding vulnerabilities in classified infrastructure are exactly what U.S. defenders need. Locking them down doesn't make the vulnerabilities disappear—it just means American defenders lose a tool while foreign adversaries continue developing equivalent capabilities without restriction. According to investingLive's report, those critics specifically warned the move could help U.S. adversaries more than it hurts them.

If the point of Project Glasswing was to find weaknesses before adversaries do, shutting down the model that found them cuts off the very feedback loop the program was designed to create.

What the Administration's Position Implies

The counterargument the administration would make, and which has not been publicly rebutted, is that a model capable of identifying classified system vulnerabilities in hours is too dangerous to leave in wide circulation, regardless of who is nominally running it. The concern isn't just about Anthropic's intentions. It's about what happens if the model or its weights reach the wrong hands. The voluntary vetting framework Trump signed gives the government up to a month to review advanced models before release. Applied to Mythos 5, that clock was apparently already running when the directive hit.

Neither the NSA nor the White House has publicly explained the specific threat calculation that drove the foreign-national restriction rather than, say, a classified access tier or a controlled research program.

What Remains Unresolved

No charges have been filed against Anthropic. No investigation has been announced. The company cooperated with Project Glasswing voluntarily, and the government's directive is a regulatory action, not a law enforcement one.

What remains unclear, as of June 24, is whether the administration's directive accomplishes anything concrete. The vulnerabilities Mythos found in classified systems still exist. The directive doesn't patch them. Project Glasswing's stated purpose was to find those gaps so they could be fixed, and it's uncertain whether the remediation work those findings were supposed to trigger is continuing, paused, or being handed off to a different tool.