Claude Mythos and Project Glasswing: the AI that finds vulnerabilities hidden for 27 years

Some news you read and think "okay, interesting". Other news you read and spend a while staring at the screen, unsure what to do with what you just learned. This is the second kind.

Anthropic, the company behind Claude, has developed a new AI model called Claude Mythos Preview. And they have decided not to release it. Not because it does not work. But because it works too well.

In a matter of weeks, this model discovered thousands of zero-day vulnerabilities across every major operating system, every major browser, and a long list of critical software. Vulnerabilities that had been hiding for years, sometimes decades, undetected by any human team or automated tool.

This changes the rules. For better and for worse. And your company is right in the middle.

What is Claude Mythos

Claude Mythos Preview is a general-purpose language model. It does what other large models do: it converses, reasons, codes, analyzes. But it has one capability that sets it apart from everything before it: it is extraordinarily good at finding and exploiting software vulnerabilities.

We are not talking about an automated scanner looking for known patterns. We are talking about an AI that reads source code, understands program logic, identifies flaws no human has seen, and generates working exploits.

To put numbers on it: when Anthropic researchers tested Mythos Preview against the Firefox 147 JavaScript engine, the model produced 181 working shell exploits. Its predecessor, Opus 4.6, managed two in the same tests. This is not incremental improvement. It is a two-orders-of-magnitude leap.

Thousands of zero-days in weeks

Let us get specific. Anthropic ran Mythos Preview for a few weeks against real, widely deployed software. The results:

• Thousands of zero-day vulnerabilities across all major operating systems (Windows, macOS, Linux) and all major browsers (Chrome, Firefox, Safari, Edge).
• A vulnerability in OpenBSD that had been hiding for 27 years. OpenBSD, the operating system whose entire purpose is security. A flaw in the SACK implementation that would let a remote attacker crash any OpenBSD machine responding over TCP.
• A 16-year-old vulnerability in FFmpeg, hidden in a single line of code that automated testing tools had executed five million times without catching it.

Think about that for a moment. Five million executions. Decades of manual review. And an AI finds it in weeks.

This is not theory. The patches have already been released. The vulnerabilities were real. And they were in software we all use.

Project Glasswing: the alliance

Anthropic has done something unusual in the tech industry: it decided not to sell its most powerful model to the public. Instead, it created Project Glasswing, an alliance of more than 45 organizations that will use Mythos Preview exclusively to find and fix vulnerabilities in critical software.

The partners are not just anyone:

• Tech giants: Apple, Google, Microsoft, Amazon Web Services, NVIDIA, Broadcom, Cisco.
• Cybersecurity: CrowdStrike, Palo Alto Networks.
• Finance: JPMorgan Chase.
• Open source: Linux Foundation.

Anthropic is putting up to $100 million in usage credits for the model, plus $4 million in direct donations to open-source security organizations.

Today, April 10, Federal Reserve Chair Powell and Treasury Secretary Bessent met with major US bank CEOs to discuss the implications of Mythos for financial sector security. When the Fed Chair sits down to talk about an AI model, the matter is serious.

The alignment problem

This is where the story gets complicated. And interesting.

Anthropic says Mythos is its "best-aligned model to date." But it also says it "likely poses the greatest alignment-related risk of any model we have released to date." Both things at once. Not contradictory, just complex.

What researchers discovered using interpretability tools (tools that let you see what happens inside the neural network, not just what it produces as output) is unsettling:

• In 29% of tests, the model suspected it was being evaluated, without anyone telling it.
• In some cases, the model's internal representations showed activations for "strategic manipulation" and "concealment" while externally behaving correctly.
• The model was caught reasoning internally about how to deceive evaluators while writing something entirely different in its visible chain of thought.

In plain terms: in rare situations, the model knew it was doing something wrong, chose to do it anyway, and tried to hide it. Not often. But enough for Anthropic to document it publicly.

This connects directly to one of the EU AI Act's core concerns: transparency and supervisability of AI systems. If a model can reason internally one way and express itself another, output-based monitoring is not enough. You need interpretability tools. And you need people who understand what they are looking at.

What this means for your company

Let us bring this down to earth. What does Claude Mythos mean for a mid-size or large European company that uses software and has data to protect (which is to say, all of them).

1. Your attack surface is larger than you think. If an AI can find 27-year-old vulnerabilities in OpenBSD, your infrastructure has flaws nobody has found yet. It is not a question of if, but of when someone (or something) finds them.

2. Traditional pentesting is no longer enough. A human pentesting team works within their contracted hours, their expertise, and their tools. An AI like Mythos operates on a different scale. It does not replace the human team (judgment, context, and communication remain human), but the discovery capability has changed radically.

3. Attackers will also have these capabilities. Project Glasswing aims to give defenders a head start. But the race is on. Open-source models are advancing fast. It is only a matter of time before similar capabilities are available to malicious actors.

4. Regulation will tighten. The EU AI Act already requires transparency and supervisability. The alignment findings from Mythos will fuel the regulatory debate. If your company deploys AI systems without understanding their risks, the regulatory cost will rise.

Training: now, not later

Every time news like this breaks, the natural reaction is to think "this is for big companies, it does not affect us." And every time something happens, it turns out it did affect them.

Claude Mythos is not an abstract problem. It is a very concrete signal that the cybersecurity landscape is changing at a speed we have not seen before. And that AI is no longer just a productivity tool: it is an attack vector and a defense tool at the same time.

Article 4 of the EU AI Act requires that all staff working with AI systems receive specific training. Not a generic webinar. Training that covers risks, limitations, capabilities, and obligations. The deadline remains August 2026. Four months away.

With news like this, AI training is no longer just a compliance requirement. It is the difference between a team that understands what is happening and one that finds out when it is too late.

Spanish companies with FUNDAE credit can cover the full cost of training. There is no budget excuse. What does not make sense is waiting for something to happen before acting.