Anthropic’s Mythos Just Dragged Wall Street Into the AI Security Race

A model release turned into a banking issue in less than a week

Anthropic published Claude Mythos Preview on April 7, 2026, and the release immediately broke out of the usual AI-news bubble. This was not another benchmark parade. According to Reuters and the Financial Times, U.S. Treasury Secretary Scott Bessent called major Wall Street banks in for a discussion about Mythos’s cyber risk potential, while British regulators prepared their own briefings for banks, insurers, market operators, and exchanges. That is a strange path for any language model. Most models appear first in GitHub repos or chat apps, then drift into developer tools. Mythos reached treasuries and central banks.

The timing matters. TechCrunch reported on April 12 that Jerome Powell joined Bessent in urging bank executives to test the model for defensive work. JPMorgan Chase appeared among Anthropic’s initial partners, and Bloomberg reporting cited Goldman Sachs, Citigroup, Bank of America, and Morgan Stanley as additional testers. A frontier model moving from a San Francisco lab into live conversations with systemically important banks in five days tells a blunt story: AI security has stopped being a lab-only topic.

Why Anthropic said no to a broad public launch

Anthropic did release details. It did not release broad access. In its April 7 technical note, the company said more than 99% of the vulnerabilities found by Mythos had not yet been patched, which is why the details stayed private. The company routed access through Project Glasswing, a limited program for critical industry partners and open-source developers. That decision turned Mythos into two stories at once: a security warning and a sales strategy.

Anthropic’s own examples were hard to ignore. The company said Mythos identified and exploited zero-day vulnerabilities across every major operating system and every major web browser during testing. It also said the oldest patched bug disclosed so far was a 27-year-old flaw in OpenBSD. In one example, Anthropic said the model wrote a browser exploit that chained four vulnerabilities, escaped a renderer sandbox, and then broke out of the operating system sandbox too. In another, it described a remote-code-execution exploit on FreeBSD’s NFS server that could grant full root access to unauthenticated users.

Those claims explain why governments paid attention. They also explain why critics smelled theater.

The numbers behind the hype, and the reason critics pushed back

Anthropic compared Mythos with Opus 4.6 on the same Firefox benchmark. Opus 4.6 reportedly produced working exploits only two times in several hundred attempts. Mythos produced 181 working exploits and reached register control 29 more times. On Anthropic’s OSS-Fuzz-style internal testing, the company said Sonnet 4.6 and Opus 4.6 each managed only one tier-3 crash, while Mythos reached full control-flow hijack on 10 fully patched targets. Those are specific numbers, and they are the reason this story dominated the past 24 hours.

But the pushback came fast. Tom’s Hardware noted that Anthropic’s headline claim about “thousands” of severe zero-days rested on 198 manual reviews. The Guardian quoted AI Now Institute chief AI scientist Heidy Khlaaf, who said the model’s capacities were not “substantiated.” Offensive-security expert Jameison O’Reilly told the paper that Mythos was a real development, yet also argued that zero-day discovery matters less in many real intrusions than people assume. His point was simple: across more than 10 years of authorized access work, the number of cases that truly required a zero-day was “vanishingly small.”

That split is the real story. Mythos may be exactly as strong as Anthropic says on exploit generation and still be oversold as a map to instant cyber catastrophe.

Why this matters more for developers than for AI gossip addicts

The easy version of the Mythos story is “new model finds bugs.” The harder version is what that means for software teams shipping code in C or C++, inside browsers, kernels, networking stacks, and infrastructure that still carries ancient assumptions. Anthropic said many of the bugs Mythos found were 10 to 20 years old. One OpenBSD bug had survived for 27 years. That should make every engineering manager wince.

This is where the category shifts from AI drama to AI coding reality. A model that can reliably move from source review to exploit scaffolding changes the economics of defensive work. Security teams can search more code in less time. Small open-source maintainers might get stronger triage help than they could ever afford from consultants. The ugly side is just as obvious: once models with Mythos-class capability spread beyond limited access, patch windows will shrink. A bug that once sat quietly for months could become weaponized over a weekend.

One prediction looks safer than most. By late 2026, major software vendors will market “AI-assisted secure development” the way cloud providers once marketed uptime. That label will not be optional. After Anthropic dragged bank regulators into this debate, boardrooms will start asking a much sharper question: which teams have model-assisted bug hunting already running inside the release pipeline, and which teams are still doing security reviews by hand.

Anthropic’s credibility problem did not vanish

Anthropic wants the public to see restraint. Critics see packaging. The company accidentally leaked part of Claude’s internal source code in early April, then spent the following week arguing that Mythos should stay tightly controlled because it found dangerous bugs too well. That sequence gave skeptics an opening they were never going to miss.

Gary Marcus put it bluntly in The Guardian: “Dario has far more technical chops than Sam, but seems to have graduated from the same school of hype and exaggeration.” Another unnamed tech PR told the paper, “Any other big tech firm would be ridiculed.” Those lines landed because they hit an existing anxiety in AI: every lab now talks about safety and capability in the same breath, and outsiders have learned to wonder which half of the sentence matters more.

Still, the banking response suggests this was not empty noise. Governments do not pull in major lenders because of a clever product launch alone. Reuters said British financial firms were preparing to hear briefings on the risks within a fortnight. Bessent’s meeting in Washington, if the reports hold up, means officials believed the model deserved immediate attention even before a mass rollout existed.

What to watch next

Two dates already matter. April 7 was the release of Mythos Preview and Project Glasswing. April 12 brought the Reuters and Financial Times reports about U.S. and U.K. financial officials reacting to the model. The next date is the one Anthropic has not given yet: the moment broader access arrives, or the company decides broader access is too risky to justify.

If Anthropic keeps Mythos inside a narrow partner ring for months, rivals will smell an opening and pitch “safe enough” cyber models with fewer restrictions. If it opens the gates soon, expect a scramble across bug-bounty programs, browser teams, Linux distributions, and enterprise security vendors. Either way, the old line between AI research and production security work already looks dead. Mythos did not simply trigger another debate about model power. It forced regulators, banks, and developers to stare at the same threat model at the same time.

For tool-by-tool comparisons, see our AI coding listings and the comparisons section.