Claude Mythos Access Gap Left the Fed Waiting for Months

The central bank that sounded the alarm couldn't use the tool

Anthropic launched Claude Mythos Preview in early April. The company said the model was unusually good at finding weaknesses and security vulnerabilities in software, so it did not release it publicly. Instead, it gave access to a select group of institutions through a cybersecurity program called Project Glasswing.1 One institution was not on the early list: the US Federal Reserve.

Reporting indicates the Fed went at least three months without access to Mythos. During that time, banks such as JPMorgan Chase, Goldman Sachs and Citi were using the model to find and fix flaws in their own systems.3 The Fed was reportedly still trying to get access as of July 15, and it is unclear whether it has received it since.1 The gap is more striking because the Fed had convened bank CEOs specifically to warn them about the model's risks.3 The central bank flagged the danger to the industry and then could not use the same tool on its own infrastructure.

How Glasswing grew, and who went first

At launch, Anthropic said roughly 50 organizations had access to Mythos. It named only a few, including JPMorgan Chase, Amazon, Apple and Google.1 Other accounts put the starting figure closer to 40.3 In June, Anthropic expanded Glasswing across 15 countries.13 The reports describe that expansion differently. One says more than 150 organizations were added in June.1 The other says the program grew to over 150 organizations in total.3 Under either count, the program grew quickly, and the Fed appears to have stayed outside it through that period.

The public record does not say why. It has not been established whether the delay came from Anthropic's own triage, regulatory or export-related constraints, procurement friction on the Fed's side, or some combination. Without that detail, the clearest conclusion is about structure. A private company decided who received frontier defensive tooling first, and that order did not match how systemically important each institution is. The Fed is arguably the most systemically important financial institution in the world, yet commercial banks were patching their systems while it waited.1

Is Mythos the real story?

Security researcher Bruce Schneier gives a more measured view of what Mythos represents. He agrees that the newest models show real gains in offensive capability. They can turn discovered vulnerabilities into working exploits without human help. They can chain several memory-corruption bugs into more complex attacks. And they can do more from a single prompt without elaborate agent setups.2

Schneier also argues that the problem is not unique to Mythos. He notes that the security firm Aisle reproduced the vulnerabilities Anthropic reported using older, cheaper, publicly available models.2 If that holds up, gating Mythos does less to keep these capabilities away from attackers than Anthropic's framing implies. Schneier suggests good PR is part of why Mythos gets so much attention.2

This point matters for how to read the Fed episode. If comparable vulnerability discovery is already possible with public models, the Fed's months without Mythos may have left it less exposed than the headlines suggest. A capable security team could have done some of the same work with other tools. The governance problem remains, though. Glasswing's premise is that defenders get a head start before attackers catch up. If an institution is left out of that head start, it falls behind the organizations that were let in, whether or not attackers have the same tool.

What it means for controlled AI rollouts

JPMorgan's Jamie Dimon has described Mythos-related risk as something governments should prioritize.3 The Fed's experience shows how far current practice is from that. Controlled release is now a standard way to ship powerful AI models, and it is often presented as the responsible option. It also makes the vendor the gatekeeper for defensive readiness across entire sectors, deciding which banks, cloud providers and countries move first.3

For enterprise AI and developer tools, the takeaway is practical. Restricting a model to a vetted group does not distribute risk evenly. It moves risk onto whoever is left out. That trade-off is easier to accept when the excluded parties are small. It is much harder to defend when one of them is the institution responsible for the stability of the financial system.

Gated rollouts will likely continue as models grow more capable. The Mythos case suggests they need clearer and more transparent rules about who qualifies, and about how critical public infrastructure gets access early rather than months later. Schneier's critique adds a caveat. If equivalent capability is already available in public models, gating may matter less than broad, fast patching across the whole ecosystem, which no single company's access list can deliver.2