Who Decides What Happens Next with the Most Dangerous Software Ever Built?
Anthropic built a model that experts worry could topple companies, industries, maybe whole economies. The government’s response? Let industry figure it out.
Last week, Federal Reserve Chair Jerome Powell and Treasury Secretary Scott Bessent yanked the CEOs of America’s largest banks out of a dinner into an emergency meeting. Jamie Dimon of JPMorgan couldn’t make it. But the heads of Goldman Sachs, Citigroup, Morgan Stanley, Bank of America, and Wells Fargo were there. The subject was a software tool that Anthropic had announced that morning — a model called Claude Mythos Preview — and what it might mean for the financial system.
What it meant, according to experts, was the potential for absolute disaster. And the meeting happened. Government representatives convened it. And then, as far as the public record shows, nothing came of it. No legislation. No executive order. No emergency regulatory framework. The banks went back to their offices, and Anthropic’s private consortium of forty companies went back to work with the tool.
This is the governance story of Claude Mythos. And so far, it’s not reassuring.
Here’s what Mythos did: over a period of weeks, according to Anthropic, it autonomously identified thousands of previously unknown security vulnerabilities — zero-days, in the parlance — across every major operating system and every major web browser. Some of these vulnerabilities had been sitting undetected for decades, surviving millions of automated tests and years of human review. In one case, a 27-year-old flaw in OpenBSD, an operating system used specifically because of its reputation for security hardiness. In another, a 16-year-old bug in FFmpeg, one of the most heavily tested media libraries in the world.
Mythos didn’t just find these flaws. It did its creepy overzealous-intern thing and figured out how to use them. It wrote code that could chain multiple vulnerabilities together, beat them all, and do it without having to enter more than a single prompt.
Anthropic’s response to building a product like this was to not release it, which is, in the annals of tech company behavior, genuinely unusual. Maybe even admirable. Instead, the company assembled a coalition it called Project Glasswing — roughly forty organizations, including Apple, Google, Microsoft, Amazon, and JPMorgan Chase — gave them early access to the model, and tasked them with patching vulnerabilities before anything comparable lands in the wrong hands. Anthropic committed $100 million in usage credits to support the effort and set a 135-day window for disclosing patched vulnerabilities. Okay, that’s good, I guess.
The security community’s reaction split roughly in half. People like Casey Ellis, the founder of Bugcrowd and one of the more clear-eyed observers of the offense-defense gap in cybersecurity, described Mythos as confirmation of something the industry already knew was coming: AI has “taken the knob that used to go to eleven and turned it to seven hundred.” The Cloud Security Alliance assembled a draft response document signed by dozens of the most prominent CISOs in the world — from Google, Cisco, Goldman Sachs, the NSA, CISA — describing the Mythos announcement as the beginning of an “AI vulnerability storm” that will “exceed anything we have experienced before.” Bruce Schneier, the security expert who is otherwise generally skeptical of AI hype, called Project Glasswing a meaningful step while noting that it remains fundamentally a reactive posture — patching holes while the surface keeps expanding.
The typical tech loudmouths were more dismissive. Famously get-out-of-our-way venture capitalist Marc Andreessen posted that the state of cybersecurity “has been dismal forever” and that perhaps now we finally have “the catalyst and the tools to fix it all.” White House AI czar and Silicon Valley impresario David Sacks wrote that Anthropic “has a history of scare tactics.” AI critic Gary Marcus raised questions about the methodology.
But none of these people are responsible for bank records, or hospital IT, or keeping the traffic lights working. And that’s where I begin to think the response to Mythos is nowhere near what history suggests it should be.
