The Most Dangerous AI Ever Built Just Found 23,000 Bugs Nobody Knew Existed
Project Glasswing Anthropic is the AI story most people haven’t read yet — and it may be the most important one of the entire year.
In April 2026, Anthropic quietly released its most powerful AI model to date — not to the public, not to developers, not even to most businesses. It released it only to a small group of the world’s most critical technology companies: Apple, Google, Microsoft, Amazon, Cisco, NVIDIA, JPMorgan Chase, Cloudflare, and roughly 40 other vetted organisations.
The model is called Claude Mythos Preview. The programme is called Project Glasswing. And in its first month of operation, it scanned over 1,000 open-source software projects and found 23,019 real vulnerabilities — bugs hidden inside software used by billions of people every single day.
One of those bugs had survived 27 years of expert human review without being caught. Another would have allowed attackers to forge security certificates across billions of IoT and industrial devices worldwide.
Less than 1% of the vulnerabilities found have been patched.
Here is the full story of Project Glasswing Anthropic — what it is, what it found, and why it should change how you think about AI.
📩 Get the AI security stories that actually matter: Subscribe to InternetBuzzFeed free → Daily AI news in plain English. No jargon.
What Is Project Glasswing Anthropic — The Full Story
Project Glasswing Anthropic was officially announced on April 7, 2026. The name comes from the glasswing butterfly — an insect with transparent wings that allow it to hide in plain sight. The metaphor is deliberate: the initiative uses AI to find the vulnerabilities hiding in plain sight inside the world’s most critical software.
The programme has two components that work together:
- Claude Mythos Preview — Anthropic’s most capable model ever built, specifically designed for advanced coding and agentic tasks. It can autonomously scan software systems, identify security flaws, understand their severity, and generate proof-of-concept exploits to confirm they are real. Anthropic considers the model so powerful that it has not released it to the general public due to cybersecurity concerns.
- A vetted partner coalition — approximately 50 organisations responsible for the software infrastructure billions of people depend on. These partners use Mythos exclusively for defensive security work — finding bugs before attacckers do.
Anthropic has committed $100 million in free model usage credits to Project Glasswing partners. The UK AI Security Institute is involved on the evaluation side. The US government has been briefed at both CISA and the Commerce Department. This is not a pilot programme. It is an emergency response to a threat Anthropic believes is already here.
| Sr. No. | Company / Organization |
|---|---|
| 1 | AWS |
| 2 | Apple |
| 3 | |
| 4 | Microsoft |
| 5 | Cisco |
| 6 | NVIDIA |
| 7 | JPMorgan Chase |
| 8 | Cloudflare |
| 9 | CrowdStrike |
| 10 | Mozilla |
| 11 | Palo Alto Networks |
| 12 | IBM |
| 13 | Linux Foundation |
| 14 | Broadcom |
Anthropic has committed $100 million in free model usage credits to Project Glasswing partners. The UK AI Security Institute is involved on the evaluation side. The US government has been briefed at both CISA and the Commerce Department. This is not a pilot programme. It is an emergency response to a threat Anthropic believes is already here.
The Project Glasswing Anthropic Results — Every Number That Matters
On May 22, 2026, Anthropic published the first month of Project Glasswing Anthropic results. The numbers are extraordinary.
To understand the scale: Mozilla patched 271 vulnerabilities in a single Firefox release (Firefox 150) based on Mythos findings. Cloudflare found 2,000 critical vulnerabilities across its infrastructure — the systems that handle traffic for a significant portion of the entire internet. A banking partner used Mythos findings to stop a $1.5 million fraudulent wire transfer mid-execution.
And then there is CVE-2026-5194 — a vulnerability in wolfSSL, a security library used in billions of IoT and industrial devices. This single bug would have allowed attackers to forge TLS certificates — the digital padlocks that prove a website or device is who it claims to be — across an estimated two billion devices globally. It had survived expert security review for years before Mythos found it.
The Bug That Survived 27 Years of Human Review
Of all the Project Glasswing Anthropic findings, one detail has dominated security community discussions more than any other: Claude Mythos found a software vulnerability that had been present in widely used code for 27 years — completely invisible to the thousands of expert human reviewers who had examined that codebase over nearly three decades.
This is not a minor edge case. It illustrates a fundamental limitation of human security review that the entire industry has known about but never fully confronted: humans get tired, humans miss patterns, humans can only hold so much context in their heads at once. A frontier AI model has none of these limitations. It can read an entire large codebase simultaneously, hold millions of lines of context at once, and apply security reasoning that draws on patterns from every vulnerability it has ever been trained on.
The 27-year bug is now patched. But the question it raises is uncomfortable: how many others like it are still out there?
Why Anthropic Is Not Releasing Claude Mythos to the Public
This is the question most people ask when they first hear about Project Glasswing Anthropic — and the answer is both reassuring and alarming.
Anthropic believes Claude Mythos Preview is the most capable AI model ever built for cybersecurity tasks. It can autonomously find zero-day vulnerabilities AND create working exploits for them. In defensive hands — like Project Glasswing partners — that capability makes the internet safer. In offensive hands, it would be one of the most powerful cyberattack tools ever created.
Anthropic explicitly stated it will not release Mythos-class models to the general public until it has developed “stronger safeguards.” Meanwhile, the EU AI Act’s next enforcement phase does not take effect until August 2026 — meaning there is a regulatory gap right now where models of this power exist with no legal framework governing their use outside Glasswing’s voluntary restrictions.
For context: when OpenAI launched GPT-5.3-Codex in February 2026, it classified the model as “high capability for cybersecurity tasks” under its Preparedness Framework — the equivalent of a warning label. That was a disclosure, not a restriction. Anthropic went further and actually restricted access. The difference in approach reflects a genuine philosophical divide about how to handle AI models powerful enough to reshape national security.
IBM Joins — And the Korean Government Gets Briefed
Project Glasswing Anthropic has been expanding rapidly since launch. Two significant developments happened this month alone.
On May 19, 2026, IBM officially joined Project Glasswing. IBM Research’s specific role: identifying and remediating vulnerabilities in widely used software and sharing findings with the broader security community. IBM becomes one of approximately 50 Glasswing launch partners — adding one of the world’s largest enterprise technology footprints to the initiative’s scanning coverage.
Separately, Anthropic opened a Seoul office in May 2026 — and within weeks, Claude Mythos Preview was being used in cybersecurity cooperation discussions with the South Korean government. The agencies involved include the Ministry of Science and ICT, National Intelligence Service, Financial Services Commission, and the National AI Strategy Committee. This is not a commercial deployment — it is government-level security briefings powered by the world’s most restricted AI model.
Project Glasswing Anthropic is quietly becoming the most significant AI security programme in the world — one government at a time.
What Project Glasswing Means for Regular People
You do not need to be a developer or security professional for Project Glasswing Anthropic to matter to you. Here is the plain-English version of why it affects everyone.
- Your browser: Mozilla patched 271 Firefox bugs found by Mythos. Every Firefox update you install from 2026 onward will likely contain Glasswing-sourced fixes.
- Your internet connection: Cloudflare handles traffic for roughly 20% of all websites. The 2,000 critical vulnerabilities found there protected traffic belonging to hundreds of millions of users.
- Your smart devices: The wolfSSL vulnerability affected billions of IoT devices — smart home devices, industrial equipment, connected cars. That bug is now patched.
- Your bank: A Glasswing banking partner stopped a $1.5 million fraudulent transfer using Mythos findings. Financial security AI is already operating in production.
- Your data: The open-source software libraries that underpin virtually every app you use are being scanned. The bugs Mythos finds today are the breaches that don’t happen tomorrow.
Final Verdict — Is Project Glasswing Anthropic Working?
The honest answer: yes, dramatically — and the numbers prove it. 23,019 real vulnerabilities found in one month, confirmed at 90.6% accuracy, across the world’s most critical software systems. By any measure, that is an extraordinary result.
The uncomfortable caveat: less than 1% patched. The bottleneck has shifted from finding bugs to fixing them. The security industry does not have enough engineers to process what Mythos is generating. Anthropic found the vulnerabilities faster than humanity can respond to them — which creates its own new category of risk.
Project Glasswing Anthropic is simultaneously the most hopeful and the most sobering AI story of 2026. The most hopeful because it proves AI can protect infrastructure at a scale and speed humans never could alone. The most sobering because it reveals exactly how fragile that infrastructure already was — and how much work remains before the vulnerabilities AI finds are actually gone.
Follow InternetBuzzFeed for every major AI security story: Subscribe free → | Read: Claude Opus 4.8 full review → | Read: Anthropic $900B valuation — why it happened → | Read: KPMG & Big Four deploy Claude →
Frequently Asked Questions
What is Project Glasswing Anthropic?
Project Glasswing is Anthropic’s restricted cybersecurity initiative, launched April 7, 2026. It gives approximately 50 vetted organisations — including Apple, Google, Microsoft, Amazon, IBM, Cloudflare, and JPMorgan Chase — access to Claude Mythos Preview, Anthropic’s most powerful AI model, for defensive security work only. The initiative is backed by $100 million in free model usage credits from Anthropic.
What did Project Glasswing find?
In its first month of results (published May 22, 2026), Project Glasswing found 23,019 vulnerabilities across more than 1,000 open-source software projects. Of those, 6,202 were classified as high or critical severity. Independent verification of a sample confirmed 90.6% were real bugs. Notable finds include 271 Firefox vulnerabilities, 2,000 Cloudflare infrastructure bugs, and CVE-2026-5194 — a wolfSSL flaw affecting billions of IoT devices.
What is Claude Mythos Preview?
Claude Mythos Preview is Anthropic’s most powerful and most restricted AI model. It is specifically capable of autonomously finding zero-day software vulnerabilities and creating working exploits. Anthropic has not released it to the general public due to cybersecurity risks, and plans to make Mythos-class models generally available only after developing stronger safeguards.
Why is less than 1% of the bugs patched?
Claude Mythos found vulnerabilities significantly faster than human engineering teams can fix them. The bottleneck has shifted from finding bugs to having enough skilled security engineers to write and deploy patches. This is now the primary challenge facing Project Glasswing — the AI outpaces the human response capacity.
Who are the Project Glasswing partners?
Confirmed Project Glasswing partners include AWS, Apple, Google, Microsoft, Cisco, NVIDIA, JPMorgan Chase, Cloudflare, CrowdStrike, Mozilla, Palo Alto Networks, IBM (joined May 19, 2026), Broadcom, and the Linux Foundation, plus approximately 40 additional vetted organisations. The UK AI Security Institute is involved on evaluation. The US government has been briefed at CISA and the Commerce Department.
Can I access Claude Mythos Preview?
Not publicly. Anthropic has restricted Claude Mythos Preview to vetted Project Glasswing partners only. It is not available through the Claude API, consumer plans, or any standard access channel. Anthropic plans to release Mythos-class capabilities more broadly only after developing safeguards sufficient to prevent misuse for offensive cyberattacks.
Visit for more infor : internetbuzzfeed.com