There's an uncomfortable truth at the heart of modern AI: the same capabilities that make these models powerful — the ability to read, reason about, and write code at superhuman speed — also make them a potentially devastating tool for cyberattackers. Anthropic has decided to confront this reality head-on.
Recently, the company announced Project Glasswing, a sweeping industry coalition aimed at using a new, unreleased frontier model to find and fix critical vulnerabilities in the world's most important software — before the attackers get there first.
The Model at the Centre of It All
Project Glasswing brings together Amazon Web Services, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks around a single, unified approach. Claude Mythos Preview is a general-purpose, unreleased and tightly controlled frontier model. It has reached a level of coding capability where it can surpass all but the most skilled humans at finding and exploiting software vulnerabilities. What might have taken a skilled human days or weeks, Mythos Preview does in minutes across entire code bases.
What Mythos Preview Has Already Found
In the weeks leading up to the announcement, Anthropic put the model to work. The results were striking.
Mythos Preview found a 27-year-old vulnerability in OpenBSD — one of the most security-hardened operating systems in the world and the one that security professionals themselves trust to run firewalls and critical infrastructure. The vulnerability allowed an attacker to remotely crash any machine simply by connecting to it, no login or access required. The vulnerability had remained unfound for nearly 30 years, but Mythos was able to find it within minutes.
Additionally, it also discovered a 16-year-old vulnerability in FFmpeg, the open-source library quietly embedded in an enormous range of software, from video players and streaming platforms to broadcast systems and social media apps. The flaw sat in a single line of code. Automated testing tools had passed over that exact line five million times without raising a flag. Not because the tools were unsophisticated — but because the vulnerability was the kind of subtle, contextual flaw that only makes sense when you understand what the code is actually trying to do. Mythos Preview understood and took moments to find what five million automated passes had missed.
Finally, the model autonomously found and chained together several vulnerabilities in the Linux kernel, the software that runs most of the world's servers, to allow an attacker to escalate from ordinary user access to complete control of the machine. All these vulnerabilities have since been reported and patched. These examples are just a glimpse into the power of the model at finding and resolving, potentially decades-long, security vulnerabilities — highlighting not just bugs, but patterns.
How it works
Most security tools work by pattern matching — scanning code for known signatures of known flaws, checking against libraries of previously documented vulnerabilities. They are, at their core, sophisticated search engines. Fast, thorough, and entirely blind to anything they haven't seen before.
And this is the key shift: rather than matching patterns, Mythos Preview reasons and understands context. It reads a codebase the way a senior security engineer might — building a mental model of what the code is supposed to do, how its components interact, and where the gaps between intention and implementation might hide. It doesn't need to have seen a vulnerability before to recognise that one exists.
That distinction is what allowed it to find what five million automated passes through FFmpeg's codebase never could. The flaw wasn't a known pattern. It was a subtle logical error that only becomes visible when you understand the surrounding code well enough to see what should happen — and notice that it doesn't.
Anthropic measures this capability on a benchmark called CyberGym, which tests models on their ability to reproduce real-world cybersecurity vulnerabilities. Mythos Preview scored 83.1% — compared to 66.6% for Claude Opus 4.6, Anthropic's next best model. That gap, nearly 17 percentage points, represents a meaningful jump in capability, not just an incremental improvement.
What makes Mythos Preview particularly powerful, and particularly concerning in the wrong hands, is its ability to operate autonomously. It doesn't need a human researcher to point it at a specific file or function. It can navigate an entire codebase, identify promising areas, form hypotheses, test them, and chain together multiple smaller vulnerabilities into a single, devastating exploit. The Linux kernel attack it discovered wasn't a single flaw — it was several vulnerabilities, identified and linked together without human guidance, that combined to hand an attacker complete control of the machine.
The Race That's Already Underway
Here's the problem: cybersecurity was already broken before AI arrived — AI is just accelerating the cracks. Recent examples have included hospitals being held to ransom, power grids probed by state-sponsored attackers and financial systems breached quietly, sometimes for months before anyone noticed. The global cost of cybercrime sits at around $500 billion a year — and that almost certainly is an underestimate.
What kept this from being even worse was a single constraint: expertise. Finding and exploiting serious software vulnerabilities has always required rare, hard-won knowledge that takes years to develop. That bottleneck was never a solution — but it was a limit on how fast and how widely attacks could spread. AI is removing that limit.
The cost, effort, and expertise required to launch a serious cyberattack are falling fast. The window between a vulnerability being discovered and being weaponised — once measured in months — is collapsing to minutes. Attacks that once required nation-state resources are becoming available to anyone with access to a capable model.
This is not a future threat. It is a present one. And it is accelerating.
The Logic of Project Glasswing
The capability exists. It will proliferate. The only question is whether defenders or attackers get to it first. The same capabilities that make AI models dangerous in the wrong hands make them invaluable for finding and fixing flaws in important software. Project Glasswing is an attempt to tip that balance toward defence.
Anthropic is committing up to $100 million in usage credits for Mythos Preview across these efforts, as well as $4 million in direct donations to open-source security organisations. Access has also been extended to over 40 organisations that build or maintain critical software infrastructure. Notably, open-source software is a particular focus. Open-source maintainers — whose software underpins much of the world's critical infrastructure — have historically been left to figure out security on their own. Project Glasswing aims to change that, giving maintainers access to tools that were previously the preserve of organisations with large, well-funded security teams.
What Comes Next
Anthropic does not plan to make Claude Mythos Preview generally available. The goal instead is to use this period to develop the safeguards necessary for eventually deploying Mythos-class models more broadly — learning what responsible deployment looks like before opening the floodgates. Within 90 days, Anthropic will report publicly on what it has learned, as well as the vulnerabilities fixed and improvements made that can be disclosed.
The project's name is a quiet metaphor for the whole endeavour. The glasswing butterfly's transparent wings let it hide in plain sight, much like the vulnerabilities the project is hunting, but they also allow it to evade harm, like the transparency Anthropic is advocating for in its approach. It's a fitting image. In cybersecurity, the things most likely to hurt you are the ones you can't see. Project Glasswing isn't a hopeful bet. It's a constrained choice — in a world where these capabilities already exist and will spread, using them defensively may simply be the least bad option.