Anthropic Just Launched Its Most Powerful AI Model, and It's Too Dangerous to Release Publicly

Anthropic dropped a bombshell this week. The company unveiled Project Glasswing, a massive cybersecurity initiative built around a new frontier model called Claude Mythos Preview. And the kicker? They're not releasing it to the public.

This one deserves your attention.

What Actually Happened

Anthropic built what it calls its most capable AI model ever. Claude Mythos Preview is a general-purpose model, meaning it can do all the things you'd expect from a frontier AI. But during testing, the team discovered something they didn't plan for: the model turned out to be extremely good at finding software vulnerabilities. Not just decent. Scary good.

We're talking thousands of previously unknown zero-day vulnerabilities found across every major operating system and every major web browser. Some of these bugs had been sitting in code for decades. One vulnerability in OpenBSD, for example, had been hiding for 27 years, as detailed in Anthropic's Frontier Red Team blog.

Nicholas Carlini, a researcher working on the project, said he found more bugs in a few weeks with Mythos than he had in his entire career. The model doesn't just find individual flaws either. It can chain multiple vulnerabilities together into complex exploit sequences, something that used to require highly skilled human hackers working over extended periods. NBC News covered how Anthropic's offensive cyber research lead Logan Graham confirmed the model can autonomously identify, weaponize, and chain together undisclosed vulnerabilities.

So what did Anthropic do? Instead of releasing Mythos to the public, they created Project Glasswing, a defensive coalition of major tech and security companies working together to patch vulnerabilities before bad actors get access to models with similar capabilities.

Who's Involved

The partner list reads like a who's who of tech and finance. According to Fortune, launch partners include Amazon Web Services, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks. On top of that, Anthropic extended access to over 40 additional organizations that build or maintain critical software infrastructure.

Anthropic is putting real money behind this too. The company committed up to $100 million in usage credits for Mythos Preview across these efforts, plus $4 million in direct donations to open-source security organizations.

Google is making Mythos Preview available to participants through Vertex AI. AWS has already been testing the model on its own critical codebases. CrowdStrike published a detailed breakdown of its role as a founding member, bringing endpoint visibility across a trillion daily events and over 280 tracked adversary groups into the coalition. This isn't a press release partnership. These companies are actively using the model.

Why This Is a Big Deal

Here's the real story. This isn't just about one model finding bugs. It's about what happens next.

Anthropic has been pretty direct about the timeline. The capabilities that Mythos has today will spread to other models soon. VentureBeat reported that Anthropic's Frontier Red Team Cyber Lead Newton Cheng warned that frontier AI capabilities are likely to advance substantially over just the next few months. The company even privately warned senior U.S. government officials that models like Mythos could make large-scale cyberattacks significantly more likely this year.

CBS News reported that Treasury Secretary Scott Bessent and Federal Reserve Chair Jerome Powell met with top bank CEOs this week specifically to discuss the cybersecurity risks posed by Mythos and similar AI systems. The IMF's managing director also weighed in, saying the world doesn't currently have the ability to protect the international monetary system against massive cyber risks. That's not the kind of language you hear every day.

For enterprise and government audiences, this changes the conversation around AI security entirely. We've moved from theoretical risk to demonstrated capability. A single AI model can now do what entire teams of security researchers would struggle to accomplish in months.

The Bigger Picture

There's a strategic angle here worth paying attention to.

Anthropic made this announcement on the same day it disclosed a major revenue milestone and a computing deal with Broadcom involving roughly 3.5 gigawatts of capacity running on Google's AI processors. The company is reportedly evaluating an IPO as early as October 2026. Security analyst Picus Security explored this tension, noting that Glasswing can be both strategically smart marketing and genuinely useful at the same time.

The competition angle matters too. OpenAI's GPT-5.3-Codex, released earlier this year, was the first model that OpenAI classified as high-capability for cybersecurity tasks. Now Anthropic has leapfrogged that with Mythos. Expect Google and others to respond with their own security-focused AI efforts in the coming months.

What to Watch Next

The real test for Project Glasswing isn't how many vulnerabilities Mythos can find. It's whether the patching can keep up. Right now, over 99% of the vulnerabilities Mythos has discovered haven't been patched yet. Open-source maintainers, in particular, are going to need serious support to handle the flood of findings.

Anthropic says it plans to launch new safeguards with an upcoming Claude Opus model, which would let them refine protections before eventually making Mythos-level capabilities more broadly available. As Simon Willison noted in his analysis, restricting the model to security researchers sounds necessary given the risks involved. Security professionals will be able to apply to a Cyber Verification Program for access.

This is one of those stories that touches everything: enterprise AI strategy, national security, open-source sustainability, and the competitive dynamics between the big AI labs. It's worth watching closely.

The era of AI-powered cybersecurity just got very real, very fast.