
Last week, Anthropic announced something that blew up FinTwit and TechTwit but didn’t seem to get lots of recognition in the annals of Wall Street power and influence. A new AI model, Claude Mythos Preview, had autonomously discovered thousands of critical vulnerabilities across every major operating system and browser, including a flaw that had been hiding inside OpenBSD for 27 years. 27 years! Rather than publish the model or its findings, handed the keys to a private coalition of tech giants and called it Project Glasswing. These companies were in a walled garden of access, using this tool to vet their systems, but still treating it like a top-secret weapon (which it basically is).
The press release was received warmly. What responsible stewardship of dangerous technology. Tech Twitter applauded. The security community cautiously chimed in with their reluctant support. Even JPMorgan put out a statement about it (they were one of the fifty companies given early access to this technology, taking part in Project Glasswing).
Nobody seemed to ask for more information about what the hell just happened.
Well, now that we’ve set up that punchline by telling you no one is asking what happened, we’re going to tell you anyway: Anthropic, a private company currently valued at $380 billion on the back of a $30 billion annualized revenue run rate (tripled from $9 billion just three months ago) selected twelve organizations to serve as exclusive stewards of a model capable of breaking virtually any software system on earth. The coalition includes AWS, Apple, Google, Microsoft, NVIDIA, Cisco, CrowdStrike, Broadcom, Palo Alto Networks, the Linux Foundation, and one financial institution: JPMorgan Chase.


JPMorgan is the sole bank at the table. The only financial institution with access to a model that has already demonstrated the ability to find and exploit zero-day vulnerabilities in “every major operating system and every major web browser.” The same model whose capabilities, per Anthropic’s own language, could cause “severe” fallout for “economies, public safety, and national security” if misused. That’s comforting, Dario.
This is not a criticism of JPMorgan. They made exactly the right call joining. The critique is more of the state that the corporate world finds itself now in, and trying to place blame on who it is that has created this scenario.
The financial system has spent the last decade obsessing over concentration risk. Too-big-to-fail banks. Central counterparty clearing houses. The AWS outage problem. The idea that when critical infrastructure is routed through a single node, the failure of that node becomes civilization-scale, and thus, they diversify across multiple nodes (or so they say).
Nevertheless, we now have that singular node in the operational segment that matters most: cybersecurity.
Project Glasswing positions Anthropic as the load-bearing wall of global software security. Mythos is being used to patch vulnerabilities in the systems that run payments, trading, custody, and communications for the largest financial institutions on earth. As one analyst noted bluntly: the world’s entire core tech stack is now downstream of Claude.
The European Systemic Risk Board’s Advisory Scientific Committee (not exactly known for alarmism) published a report in December warning precisely about this dynamic. Concentration among a small number of AI providers, they wrote, creates “highly interconnected nodes” in the financial system. When multiple institutions rely on the same model, their risk profiles converge. Their exposures correlate. And the failure of a single provider, whether through technical failure, security breach, or model error, produces cascading effects that no individual institution can contain.
That was a theoretical warning. Project Glasswing has made it operational.
And of course, it gets political.
The same week Anthropic announced Project Glasswing, it disclosed that it is actively considering an IPO as early as October 2026, with a target raise north of $60 billion. Goldman Sachs, JPMorgan, and Morgan Stanley are reportedly competing for underwriting roles. The company that just became the de facto security infrastructure for global financial markets is preparing to go public — with the same banks whose security it now manages potentially earning hundreds of millions in underwriting fees to make it happen.
This is less an allegation of wrongdoing and more an inquiry about structure. When the institution responsible for protecting a system has a direct financial stake in the valuation of the company that built the protection, the incentives are no longer cleanly aligned with systemic safety. This is the same logic that made analysts conflicted in the dot-com era, that made rating agencies conflicted in 2007, and that makes anyone covering their own book conflicted in every era.
There is a reason we have disclosure rules, firewall requirements, and conflict-of-interest regulations in financial services. Those rules were built on the insight that structural incentives corrupt judgment even among people of goodwill. The question is not whether Jamie Dimon is corrupt. The question is whether the architecture of Project Glasswing is sound and whether anyone with regulatory authority is actually looking at it.
All signs point to no.
Washington’s approach to AI governance in 2026 has been, broadly, to let the industry self-organize. Project Glasswing is self-organization in its purest form: a private company, facing no legal requirement and no external oversight, decided who gets access to a potentially civilization-altering capability. The decision was made by a team in San Francisco, ratified by a press release, and covered mostly as a tech story.
The European Systemic Risk Board’s report urged that regulators “ideally develop their own AI infrastructures” to avoid dependence on industry self-reporting, explicitly drawing the comparison to how supervisors failed to understand complex derivatives before 2008. American regulators have not taken that advice. The Fed, the OCC, and the FDIC are still processing what it means that their regulated institutions are now running AI models for routine operations. The notion that they are equipped to evaluate the systemic implications of a private cybersecurity consortium operating at this scale is not credible.
Meanwhile, the White House summoned representatives from JPMorgan, Goldman, Citi, Bank of America, and Morgan Stanley to discuss vulnerabilities surfaced by Mythos. That meeting happened. It was not a regulatory hearing. There were no binding outcomes. It was a briefing. The government was informed.
None of this means Project Glasswing is wrong. The logic of the initiative is correct: if a model this powerful exists, better to use it for defense before adversaries develop comparable capability. Anthropic’s choice to restrict access rather than release widely reflects genuine restraint. JPMorgan’s participation is prudent.
But “better than the alternative” is not the same as “structurally sound.” The financial system spent decades learning that private coordination among large institutions not a substitute for public oversight. MBS were well-intentioned. LIBOR was well-intentioned. The repo market was well-intentioned. The pattern is consistent: when the people managing a risk are also the people profiting from the infrastructure that creates it, the risk gets systematically underpriced until it doesn’t. The government is the primary market agent that has essentially nil incentive risk and can therefore best regulate and oversee innovations like these.

Project Glasswing has created a single point of failure for global cybersecurity, operated by a pre-IPO company, governed by a private consortium, and supervised by no one with statutory authority to do anything about it if something goes wrong. The security community is asking whether Mythos can fix more vulnerabilities than it surfaces. The IPO community is asking what the float will look like in October. Someone should be asking whether we have just privatized critical financial infrastructure under the cover of a press release and how long it’s going to take for the world to notice that our entire economy is converging on a handful of key men and key balance sheets.
That’s all for today.
Thanks for reading & stay curious folks!
-J&E