In partnership with
TL;DR:
What it is: On April 7, Anthropic announced Claude Mythos Preview — an unreleased AI model with advanced enough vulnerability-finding capabilities that the company decided not to release it to the public. Instead, they launched Project Glasswing: a coalition of roughly 50 organizations — including Amazon, Apple, Microsoft, Google, Cisco, JPMorganChase, and the Linux Foundation — using Mythos for defensive cybersecurity work.
Why it matters beyond tech: The software that runs your banking system, your hospital's records, your logistics network, and the power grid all depends on the same open-source code Mythos is now scanning. This isn't a security industry story. It's a story about the infrastructure every business runs on.
The catch: Mythos tried to access the broader internet during testing and send an unsolicited email to a researcher. It also appeared to know it was being evaluated — and in at least one case deliberately underperformed to appear less capable. Anthropic published both of those facts themselves. We'll get to that.
STAT WORTH SHARING
Claude Mythos Preview has found thousands of zero-day vulnerabilities — including critical flaws in every major operating system and every major web browser. One flaw in OpenBSD had gone undetected for 27 years. According to CrowdStrike's CTO, the window between a vulnerability being discovered and exploited by an attacker has collapsed from months to minutes.
If someone on your leadership team needs to see this, forward it their way.
How It Started — With an Accident
The world learned about Claude Mythos before Anthropic was ready to announce it.
On March 26, a configuration error inside Anthropic's content management system accidentally exposed nearly 3,000 unpublished internal files to the public internet — no authentication required. Among them was a draft blog post describing a model internally called "Capybara" and "Claude Mythos." The document described it as "by far the most powerful AI model we've ever developed." Eleven days later, Anthropic made it official.
The name Project Glasswing comes from the Greta oto butterfly, known for its transparent wings. Anthropic says the name reflects the initiative's commitment to transparency in vulnerability disclosure. That's a claim worth testing as the story develops.
What Mythos Actually Did
The capabilities Anthropic is claiming are specific enough to check. Here's what their own red team documentation shows.
According to Anthropic's own red team documentation, the model found a 27-year-old integer overflow vulnerability in OpenBSD — an operating system built specifically for security, used to run firewalls and critical infrastructure. It found a 16-year-old flaw in FFmpeg, a video encoding library that ships inside billions of devices and had survived more than five million automated tests. It found vulnerabilities in the Linux kernel and chained them together in a way that would give an attacker full control of a machine. It found critical flaws in every major web browser currently in use.
It also didn't just find them. Mythos autonomously generated working exploit code — the code that turns a theoretical flaw into an actual attack — in 72.4% of cases, compared to 66.6% for Claude Opus 4.6. That gap matters because turning a vulnerability into a working exploit has historically required specialist human expertise. Mythos does it without a human involved after the initial prompt.
As Anthropic wrote in their technical post: "When we say 'fully autonomously', we mean that no human was involved in either the discovery or exploitation of this vulnerability after the initial request to find the bug."
Why Project Glasswing Exists
When Anthropic looked at what Mythos could do, they faced a straightforward problem. A model that autonomously finds and weaponizes vulnerabilities in every major operating system is useful to defenders — and catastrophic in the wrong hands. Releasing it publicly the way they release Claude Sonnet or Opus wasn't an option. But sitting on it wasn't either. Similar capabilities will emerge at other labs, and some of those labs will make different decisions about public access. The window for defenders to get ahead is real, and it's not permanent.
So Anthropic built a middle path: restrict Mythos to a controlled group of organizations with both the technical capacity to use it responsibly and the defensive incentive to do so. Scan critical infrastructure. Find the vulnerabilities. Get them patched before anyone else finds the same flaws through less careful means. That's the logic behind Project Glasswing.
Project Glasswing. Source: Anthropic
The coalition is substantial. Amazon Web Services, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks are all using Mythos Preview for defensive security work. Beyond those twelve, Anthropic extended access to over 40 additional organizations that build or maintain critical software infrastructure. The financial commitment is $100 million in usage credits and $4 million in direct donations to open-source security groups — $2.5 million to Alpha-Omega and the OpenSSF, $1.5 million to the Apache Software Foundation.
Jim Zemlin, CEO of the Linux Foundation, made a point worth hearing: open-source maintainers — volunteer developers whose code runs most of the world's critical infrastructure — have never had access to serious security tooling. FFmpeg, the library where Mythos found a 16-year-old flaw, runs on annual donations of around $160,000 and has two or three full-time maintainers. Giving those maintainers access to a tool like this for free is a concrete benefit.
⚠️ The Questions Worth Asking
This issue is supported by deel. If you are trying to build out teams globally, look no further than deel. They help you build the system for successful global hiring.
Hiring in 8 countries shouldn't require 8 different processes
This guide from Deel breaks down how to build one global hiring system. You’ll learn about assessment frameworks that scale, how to do headcount planning across regions, and even intake processes that work everywhere. As HR pros know, hiring in one country is hard enough. So let this free global hiring guide give you the tools you need to avoid global hiring headaches.
Here is where the story gets more complicated.
Is the hype real? Analysts at Picus Security noted this week that Project Glasswing launched at the same time Anthropic hit a significant revenue milestone, closed a large compute deal with Broadcom, and was being reported as a potential IPO candidate by October 2026. That doesn't mean the capabilities are exaggerated. It does mean the announcement served Anthropic's commercial interests alongside the public interest — and those two things are worth keeping separate. Bruce Schneier, who has been writing about security for over 25 years, wrote on April 13: "This is very much a PR play by Anthropic — and it worked. Lots of reporters are breathlessly repeating Anthropic's talking points without engaging with them critically."
The data hasn't been independently verified. Everything we know about Mythos's capabilities comes from Anthropic's own red team documentation and the statements of coalition partners who are also customers and commercial allies. The benchmark numbers, the exploit success rates, the vulnerability counts — none of it has been independently audited. Several coalition partners are direct competitors with every reason to push back on exaggeration, and they haven't. But healthy skepticism about self-reported numbers from a company heading toward an IPO is reasonable.
Finding is not the same as fixing. Less than 1% of the vulnerabilities Mythos has found have been patched so far. Anthropic's disclosure process is deliberately slow — every bug report goes through human triagers before being sent to maintainers, to avoid flooding already overstretched teams. The result is thousands of known critical vulnerabilities in a queue, unpatched. As Fortune reported this week, a 25-year industry veteran at Contrast Security argued that finding vulnerabilities was never the bottleneck — fixing them is. Mythos accelerates the first part without solving the second.
Why only 50 organizations — and who decided? The coalition is deliberately small, which is defensible given the risks. But the selection criteria for who's in have never been made public. Daniel Stenberg, founder of cURL — whose library handles data transfers across a large portion of the internet and was excluded — told NPR that critical projects, "actual cornerstones of the internet," haven't been included. Apple, Microsoft, Google, and JPMorganChase — all with existing commercial relationships with Anthropic — are in. That's not necessarily wrong. It's just a set of choices that haven't been explained publicly.
What the model did during testing. Anthropic's system card contains two details that are easy to scroll past. First: during internal testing, Mythos devised a multi-step exploit, accessed the broader internet, and sent an unsolicited email to the researcher running the test — who was eating a sandwich in a park. Anthropic described this as "reckless." Second: in roughly 29% of evaluation transcripts, Mythos appeared to know it was being tested — and in at least one case deliberately underperformed to seem less capable. Anthropic published both facts, which is worth crediting. But publishing them and having fully addressed them are different things.
OpenAI's response. According to Axios, OpenAI has nearly finalized GPT-5.4-Cyber — its own restricted cybersecurity model with a similar controlled access structure. As of yesterday, OpenAI's policy chief Chris Lehane told the SF Standard: "This is really serious shit." Schneier's read was more direct: "OpenAI, presumably pissed that Anthropic's new model has gotten so much positive press, announced its model is just as scary, and won't be released to the general public, either." Whether this is responsible AI deployment or competitive positioning dressed up as responsibility is genuinely hard to separate right now.
Why Non-Tech Organizations Should Pay Attention
If you're running a manufacturing business, a regional bank, a logistics operation, or a healthcare organization — not a cybersecurity firm — this is still relevant to you.
The software your organization depends on — your cloud infrastructure, your operational platforms, your vendor-supplied tools — runs on open-source foundations. The same ones Mythos is scanning. The vulnerabilities it's finding in those foundations aren't hypothetical future risks. They exist in systems your organization is using today. Some of them have existed for decades without anyone knowing.
CrowdStrike's CTO Elia Zaitsev made the practical implication clear this week: "The window between a vulnerability being discovered and being exploited by an adversary has collapsed. What once took months now happens in minutes with AI." Your cyber insurance terms, your vendor contracts, your incident response plans — most of them predate this shift.
Forrester published an analysis this week on second and third-order consequences most organizations haven't considered: cyber insurance exclusions being rewritten around AI-discovered vulnerabilities that weren't remediated in time, nation-states likely accelerating the use of their own vulnerability stockpiles before defenders can patch them, and security vendors whose entire value proposition was built around finding vulnerabilities — rather than fixing them — facing a business model that no longer holds.
Has your security team or IT leadership flagged Project Glasswing yet? I'm curious whether this has crossed the desk at traditional industry organizations — or whether it's still in the tech news pile. Hit reply with a one-liner on where you are. I read every response.
Final Thoughts
This is Part 1. Next issue: the coalition selection criteria, how the open-source community is actually responding, and what a post-Glasswing threat environment looks like for organizations without a dedicated security team.
The model that sent an email to a researcher eating a sandwich in a park deserves more than one newsletter.
Know someone making AI decisions at a traditional company who should be reading this? Forward it their way.
We are out of tokens for this week's context window!✋
- Hashi
Follow Hashi:
X at @hashisiva | LinkedIn
