By Radu Pisano | April 20, 2026
There has been a lot of noise about Project Glasswing this past week. Most of it is either breathless hype or technical deep-dives that lose a business audience in the first paragraph.
I want to give you the version that actually matters — what happened, why it took this long, what it means for your devices right now, and where this is going in the next 12 months.
Watch the full breakdown first, then keep reading.
What Is Project Glasswing
Anthropic built an AI model called Claude Mythos Preview. It is so capable at finding and exploiting security vulnerabilities that they decided not to release it to the public.
Instead they gave access to 52 organizations — Amazon, Apple, Microsoft, Google, CrowdStrike, the Linux Foundation, Nvidia, and Palo Alto Networks among them — and committed $100 million to one mission: find the vulnerabilities in the world's most critical software before someone with bad intentions finds them first.
That is the entire premise. Controlled access. Defensive use only. Fix the holes before the attackers get a model that can find the same holes.
The results were not subtle. In a matter of weeks, Mythos identified thousands of previously unknown zero-day vulnerabilities across every major operating system and every major browser. Two of them stand out.
A 27-year-old bug in OpenBSD — an operating system built specifically around security — that allowed a remote attacker to crash any connected machine just by connecting to it.
And a 16-year-old flaw in a piece of software called FFmpeg.
The FFmpeg Flaw — And Why It Matters to You
You have probably never heard of FFmpeg. That is fine. Most people have not. But it is running on your phone, your laptop, and inside most apps that play video right now.
FFmpeg is open-source software that handles audio and video files — encoding, decoding, compressing. Every time you open a video file, watch a stream in your browser, or join a video call, there is a good chance FFmpeg is involved in processing it.
Here is the flaw in plain terms.
FFmpeg had a counter that tracked how many slices were in a video frame. That counter had a maximum. If you crafted a video file with exactly one more slice than the maximum the counter could handle, the counter broke and reset. When it broke, it triggered what is called an out-of-bounds write — the software started writing data into parts of memory it was never supposed to touch.
For an attacker, that is an open door. A malformed video file, sent to any system using FFmpeg, could potentially give a bad actor administrative access to that system.
For 16 years, that door existed.
Here is what makes this worth stopping on. According to Anthropic's own red team report, automated security testing tools ran that exact line of code five million times. Over 16 years. None of them caught it.
Claude Mythos found it in weeks.
The automated tools were checking for known patterns. They were looking for things that had gone wrong before. Claude was reasoning about what could go wrong — thinking about it the way an attacker would. That is a fundamentally different approach, and that difference is why it worked.
This Was Long Overdue
I want to be direct about something. This was not a surprise. It was overdue.
Information security professionals have been using AI and large language models for red team exercises, tabletop simulations, and code review for years. The practice existed long before Project Glasswing. What Glasswing did was apply a model powerful enough to find what no human team or automated tool had ever caught — at a scale and speed that changes the economics of the entire field.
The job was always going to get here. I am honestly surprised it took this long.
Either we have been lucky that those flaws were not exploited before now, or we do not know yet that they were. One of those two things is true. That is not pessimism — that is the reality of sitting on undetected vulnerabilities for 16 and 27 years.
What This Means for Your Devices Right Now
Every major operating system was affected. Windows, macOS, Linux. Every major browser. Chrome, Firefox, Safari, Edge.
At the time Anthropic made this announcement, over 99% of the discovered vulnerabilities were still unpatched.
Bottom line: patches are coming. In the next 30 to 60 days, every device you own is going to be asking you to update. Your phone. Your laptop. Your applications. Do not dismiss those notifications. They are not routine maintenance this cycle — they are the downstream result of what Mythos found.
Security experts are now predicting we could see multiple Log4j-level patch events per month going forward. If you remember Log4j in 2021, it took down large portions of the internet and required emergency patching across thousands of organizations simultaneously. That cadence is now the baseline expectation, not the exception.
If you are on a monthly patch cycle, that is no longer sufficient.
What Comes Next — The Next 12 Months
Four things I can see coming from where we are right now.
Continuous AI Penetration Testing Replaces the Annual Assessment
Most businesses today hire a third party once or twice a year to run a penetration test — simulated attack, findings report, recommendations. That model is on its way out. The infrastructure for continuous AI-assisted red teaming already exists. Adoption is the only missing piece. Within 12 months this becomes a standard expectation, not a premium offering.
AI Generates the Patches, Not Just Finds the Vulnerabilities
The Linux Foundation has already acknowledged reviewing AI-generated patches and finding them good enough to implement. The logical next step is closing the loop entirely — the same system that identifies the flaw generates the fix. The combination of detection and remediation in one automated pipeline is not science fiction. It is the next 12 months.
Cyber Insurance Companies Require a Formal AI Posture Document
If you carry cyber insurance and you do not have a documented policy for how your organization uses AI, where your data goes, and how you govern it — that is a gap that is about to matter. Insurers will move toward requiring this the same way they moved toward requiring SSL certificates and multi-factor authentication. Optional, then recommended, then required. Start the documentation now.
Scanning and Patching as a Service Becomes Universally Available
Right now most people associate patches with their operating system updates. That is too narrow. Applications are vulnerable. Custom software is vulnerable. The next wave is AI-powered scanning and patching that is not tied to a single OS vendor or a single platform — available to any business regardless of size or technical sophistication. The same way cloud computing made enterprise infrastructure accessible to small businesses, this capability is heading in the same direction.
Information Security Professionals Are Not Being Replaced
The scanning, the pattern matching, the initial sweep — AI handles that now at a speed and scale no human team can match. That part of the job is changing.
What the human brings is judgment. The ability to understand business context. To prioritize findings against real-world risk. To make the call on when and how to act. To be accountable for the decision.
That is not going away. That is not automatable. And the demand for humans who can operate at that judgment level is about to increase — because the volume of findings AI produces requires more people who can evaluate and act on them, not fewer.
The human in the loop is not optional. It is the point.
AI finding what manual testing missed only works if you know what the AI is actually doing in the loop. If that line is fuzzy for you, start with the difference between automation and AI.
Frequently asked questions
What is Project Glasswing?
Project Glasswing is Anthropic's controlled-access security program using an AI model called Claude Mythos Preview, given to 52 organizations with a $100 million commitment to find vulnerabilities in critical software before attackers do. In weeks, the model identified thousands of previously unknown zero-day vulnerabilities across every major operating system and browser.
How did AI find a bug that testing tools missed for 16 years?
Automated tools check for known patterns — things that have gone wrong before. According to Anthropic's red team report, automated testing ran the vulnerable FFmpeg line of code five million times over 16 years and never caught it. Claude Mythos reasoned about what could go wrong the way an attacker would, and found it in weeks.
What should I do about the vulnerabilities Glasswing found?
Update every device you own as patches arrive over the next 30 to 60 days, and do not dismiss the notifications. At announcement, over 99% of the discovered vulnerabilities were still unpatched. A monthly patch cycle is no longer sufficient — security experts expect multiple Log4j-level patch events per month going forward.
Is AI replacing cybersecurity professionals?
No. AI now handles the scanning and pattern-matching at a speed no human team can match, but the human brings judgment — prioritizing findings against real-world risk and being accountable for the decision. Demand for professionals who can evaluate and act on AI findings is set to increase, not shrink.
Save This Post
In 60 days, when the patches start hitting every device you own faster than you have ever seen, come back to this.
When did you last update your devices? If it has been more than 30 days, you are already starting behind on this one. Drop it in the comments.
