Anthropic's Most Powerful AI Model Isn't For You — It's Hunting Zero-Days Instead

News Summary

On Tuesday, April 7, 2026 (Eastern Time), Anthropic officially announced Claude Mythos Preview, a new general-purpose frontier AI model that the company describes as the most capable it has ever built. In a departure from a conventional public release, Anthropic is not making Mythos available to the general public. Instead, it has launched the model exclusively through a coordinated initiative called Project Glasswing, channeling Mythos's capabilities toward defensive cybersecurity work.

What Is Claude Mythos Preview?

Claude Mythos Preview is Anthropic's newest and most advanced language model, significantly outperforming all prior Claude models across general benchmarks. While it was not purpose-built for security, it demonstrates remarkable aptitude for agentic coding, reasoning, and vulnerability analysis. On the CyberGym benchmark — which evaluates AI agents on vulnerability analysis tasks — Mythos Preview scored 83.1%, compared to 66.6% for Claude Opus 4.6, which previously led the leaderboard. Anthropic describes the model as "far ahead of any other AI model in cyber capabilities."

The model's existence was first revealed in late March 2026 when a misconfiguration in Anthropic's content management system inadvertently published an internal draft blog post, which described Mythos as "by far the most powerful AI model" the company had developed and warned of its potential cybersecurity implications.

Project Glasswing: A Coalition for Cyber Defense

At the center of the Mythos launch is Project Glasswing, a new industry coalition designed to use AI offensively for defensive purposes — finding and patching vulnerabilities before they can be exploited. The founding partners of Project Glasswing include some of the most prominent names in global technology and cybersecurity:

Amazon Web Services, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks.

In addition to the 12 core partner organizations, approximately 40 additional organizations responsible for building or maintaining critical software infrastructure have also been granted access to the Mythos Preview. Partners will conduct defensive security assessments using the model and commit to sharing their learnings with the broader industry, helping raise the overall security baseline across the technology ecosystem.

Thousands of Zero-Day Vulnerabilities Discovered

In the weeks leading up to the announcement, Anthropic ran Mythos Preview against a wide range of software systems and reported that the model identified thousands of zero-day vulnerabilities — previously unknown security flaws. Many of these were rated as high-severity and critical, and a significant number had gone undetected for years despite extensive human and automated review.

Notably, Mythos uncovered vulnerabilities across every major operating system and web browser. Among its findings was a 27-year-old bug in OpenBSD — an operating system long considered a gold standard for security. In one demonstration, the model identified and chained together a Linux kernel flaw that could allow complete control over a machine. Most of these vulnerabilities are currently unpatched or have only been patched very recently.

Anthropic's technical blog published detailed analysis of how the model autonomously reasons about codebases, formulates hypotheses about potential vulnerabilities, runs tests, and produces proof-of-concept exploits — all without human guidance beyond an initial prompt.

Why Not Release It Publicly?

Anthropic has been explicit about its reasoning: the same capabilities that make Mythos Preview so effective at finding and fixing vulnerabilities could, in the wrong hands, be weaponized to discover and exploit them at scale. A draft memo that leaked before the official announcement cautioned that Mythos "presages an upcoming wave of models that can exploit vulnerabilities in ways that far outpace the efforts of defenders."

CrowdStrike CTO Elia Zaitsev echoed the concern, noting that the window between vulnerability discovery and exploitation has already collapsed — what once took months now happens in minutes with AI. In that context, Anthropic chose a carefully staged rollout to give defenders a head start before wider access is considered.

Anthropic has stated it does not plan to make Claude Mythos Preview generally available in its current form, but acknowledged that its "eventual goal is to enable users to safely deploy Mythos-class models at scale — for cybersecurity purposes, but also for the myriad other benefits that such highly capable models will bring."

Financial Commitment and Open-Source Support

To support Project Glasswing, Anthropic has committed up to $100 million in API usage credits so that partner organizations can conduct security testing at scale without bearing the infrastructure costs. In addition, the company is donating $4 million to open-source security organizations, reinforcing its commitment to the broader software security community.

Claude Mythos Preview is also available in Private Preview to select Google Cloud customers through Vertex AI, as part of the broader Project Glasswing rollout.

Industry Impact and Looking Ahead

News of Mythos's capabilities had already reverberated through financial markets before the official announcement. Following the earlier leak in March, shares of multiple cybersecurity companies — including CrowdStrike, Palo Alto Networks, Zscaler, SentinelOne, and Okta — fell between 5% and 11%, reflecting investor concerns about how frontier AI models might reshape the competitive landscape of the security industry.

Anthropic's approach with Mythos Preview represents a new model for responsible AI deployment: rather than broad public access, the company is making its most capable model available selectively to those best positioned to use it safely and constructively — while simultaneously making the case that AI must become a core pillar of modern cybersecurity infrastructure.