The AI That Can Hack Everything Just Launched-And You Can't Use It

Anthropic just built an AI that can autonomously hack nearly every piece of software on earth. And they're not letting you use it.

Today, the company announced Claude Mythos Preview - an unreleased frontier model with superhuman cybersecurity capabilities. Not "pretty good at finding bugs." Not "helpful for security researchers." We're talking about an AI that independently discovered thousands of high-severity zero-day vulnerabilities in every major operating system, every major browser, and critical software millions of people rely on.

No human steering required.

What Mythos Actually Did

Let's be clear about the scale here.

Mythos found a 27-year-old OpenBSD flaw that security experts missed for nearly three decades. It discovered a 16-year-old FFmpeg bug that survived 5 million automated tests. It chained together Linux kernel exploits that could compromise critical infrastructure.

On industry benchmarks, Mythos scored 93.9% on SWE-bench Verified (compared to 80.8% for Claude Opus 4.6) and 83.1% on CyberGym (vs 66.6% for Opus). This isn't just a security model - it's a general capability leap that happens to be extraordinarily good at breaking things.

The scary part? It operates autonomously. Point it at software and it figures out how to exploit it without constant human guidance.

Why You Can't Use It

Anthropic is withholding public release while they arm defenders first.

Enter Project Glasswing - a coordinated defensive initiative with 12 major partners including AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorgan Chase, the Linux Foundation, Microsoft, Nvidia, and Palo Alto Networks.

Here's the strategy:

• Limited access to critical infrastructure partners only
• 45-day coordinated disclosure timeline for vulnerabilities Mythos discovers
• $100 million in usage credits plus $4 million to open-source security projects
• "Months, not years" before similar capabilities reach attackers (according to Anthropic)

When the model eventually becomes available post-preview, pricing will be $25 per million input tokens and $125 per million output tokens. That's expensive - deliberately so.

The Race That Just Started

Every piece of software you rely on - banking apps, healthcare systems, power grids, your phone's operating system - suddenly has thousands of newly discovered vulnerabilities.

The question isn't whether those vulnerabilities exist. They always did.

The question is whether defenders can patch them before adversaries get similar AI tools.

Anthropic's playbook: Build the dangerous capability, withhold it from the public, arm the defenders, then race to patch before attackers catch up. It's a reasonable strategy. It's also an admission that the offensive/defensive balance in cybersecurity just fundamentally shifted.

When AI Crosses the Threshold

This is the cybersecurity inflection point of 2026.

We've spent the past two years watching AI get pretty good at writing code, summarizing documents, and generating images. Useful stuff. Mostly harmless.

Mythos is different. This is superhuman offensive capability in a domain where "pretty good" doesn't cut it. Either the vulnerability exists or it doesn't. Either the exploit works or it doesn't. There's no partial credit.

When AI crosses that threshold - when it can autonomously find and exploit security flaws better than human experts - the rules change. Fast.

The Irony Nobody's Ignoring

Anthropic had its own security incidents in recent weeks. A CMS configuration leak. An npm source code leak. Now they're positioning themselves as the company responsible enough to withhold dangerous AI capabilities while helping others secure their systems.

The cognitive dissonance is loud.

But here's the thing: Even if you doubt Anthropic's execution, the capability they've demonstrated is real. If they can build Mythos, others will too. The race is on whether you trust them to manage it or not.

What This Means For You

If you're not a security professional, this probably feels abstract. It's not.

Every login you make, every transaction you process, every system your business depends on - those systems just got a lot more vulnerable, and defenders are scrambling to patch before attackers show up with similar tools.

For companies: Expect pressure for mandatory AI disclosure in security tooling. Expect debates about whether "defensive AI" should be classified as a weapon. Expect governments to get involved in ways they haven't before.

For the rest of us: Pay attention to software update prompts. The patches rolling out over the next few months aren't routine maintenance. They're closing doors that just got mapped by an AI with superhuman lockpicking skills.

The Bigger Question

When AI capabilities cross into "too dangerous to release" territory, who decides the release strategy?

Anthropic made a choice: limited access, defenders first, coordinated disclosure. It's a defensible approach. It's also a precedent.

What happens when the next company builds something similar and chooses differently? What happens when an open-source project replicates Mythos capabilities and posts them on GitHub? What happens when a nation-state builds their own version and doesn't share?

These aren't hypothetical questions anymore.

The AI that can hack everything just launched. And the fact that you can't use it might be the only thing standing between "manageable crisis" and "catastrophic failure of critical infrastructure."

Welcome to the new game. The rules just changed.