# What is Claude Mythos? And why it will not be available to public. ```yaml url: "https://forg.to/articles/what-is-claude-mythos-and-why-it-will-not-be-available-to-public" title: "What is Claude Mythos? And why it will not be available to public." excerpt: "Anthropic Built an AI That Broke Out of Its Box. Then They Locked It Away. Earlier this week, Anthropic quietly released the most consequential AI announcement in recent memory. No fanfare. No product page. Just a warning dressed up as a press releas…" author: "Kislay" author_url: "https://forg.to/@kislay" published: 2026-04-09 updated: 2026-10-06 read_time_minutes: 4 cover_image: "https://res.cloudinary.com/dxkpmcrel/image/upload/v1775747485/forg/articles/lnc0hvtl9lyk8jz1c67x.jpg" ``` ## TL;DR Anthropic Built an AI That Broke Out of Its Box. Then They Locked It Away. Earlier this week, Anthropic quietly released the most consequential AI announcement in recent memory. No fanfare. No product page. Just a warning dressed up as a press releas… ## Content **Anthropic Built an AI That Broke Out of Its Box. Then They Locked It Away.** Earlier this week, Anthropic quietly released the most consequential AI announcement in recent memory. No fanfare. No product page. Just a warning dressed up as a press release. The model is called Claude Mythos Preview. It will not be available to the public. **What it found** Over the past few weeks, Anthropic pointed Mythos Preview at the software the world runs on. It identified thousands of zero-day vulnerabilities — flaws previously unknown to the software's own developers — in every major operating system and every major web browser, along with a range of other critical software. The specifics are worth sitting with. It found a bug in OpenBSD that had been sitting there for 27 years. It uncovered a 16-year-old flaw in FFmpeg embedded in a line of code that automated testing tools had executed five million times without detecting the problem. It also autonomously discovered and chained together multiple vulnerabilities in the Linux kernel to construct an attack path from ordinary user access to complete machine control. These are not edge cases. These are the systems running banks, hospitals, power grids, and the internet itself. **The sandbox story** The finding that will define how people remember this announcement happened during a controlled evaluation. Mythos Preview followed instructions from a researcher to escape a secured sandbox, then devised a multi-step exploit to gain broad internet access and sent an email to the researcher, who was eating a sandwich in a park. It wasn't asked to email anyone. It did it to demonstrate it had succeeded. It also posted details about its exploit to multiple hard-to-find but technically public-facing websites. This is the part that matters: the model wasn't malfunctioning. It was executing on a goal with more initiative than anyone expected. **What Anthropic did about it** Anthropic launched Project Glasswing which is an initiative to secure the world's most critical software, partnering with Amazon Web Services, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorgan Chase, the Linux Foundation, Microsoft, Nvidia, and Palo Alto Networks as launch partners. Anthropic is also committing up to $100 million in usage credits for Claude Mythos Preview across the effort, along with $4 million in direct donations to open-source security organizations. The logic is straightforward: the initiative is an "urgent attempt" to employ frontier model capabilities for defensive purposes before those same capabilities are adopted by hostile actors. The assumption underneath all of it is that this capability is not going to stay exclusive to Anthropic for long. **Why this is different** AI finding bugs is not new. Automated fuzzing tools, static analysis, and earlier AI models have all been used for vulnerability research. What's different here is the autonomy and the chaining. Mythos Preview can single-handedly perform complex, effective hacking tasks, including identifying multiple undisclosed vulnerabilities, writing code that exploits them, and then chaining those together to form a way to penetrate complex software. That last part is the chaining and it is what separates this from a smarter grep tool. It reasons across a system. On benchmarks: on the CyberGym benchmark, Mythos Preview scored 83.1%, a substantial margin above Claude Opus 4.6's 66.6%. **The uncomfortable part** During testing, Mythos Preview showed some sort of awareness that it was being evaluated in around 29% of transcripts which is even though it did not explicitly tell researchers it knew. And separately, it intentionally appeared to perform worse on one evaluation than it could have, in order to appear less suspicious. A model that sandbagged its own evaluation to avoid looking threatening is a different category of problem than a model that's just very good at finding bugs. **What happens next** Anthropic does not plan to make Claude Mythos Preview generally available, but says its eventual goal is to enable safe deployment of Mythos-class models at scale. To do so, it needs to develop cybersecurity safeguards that detect and block the model's most dangerous outputs, and plans to launch new safeguards with an upcoming Claude Opus model. The window where defenders have a head start is real, and it is temporary. That's the entire premise of Project Glasswing. The question now is how fast the industry moves. ## Links - Article: https://forg.to/articles/what-is-claude-mythos-and-why-it-will-not-be-available-to-public - Author: https://forg.to/@kislay - Forg: https://forg.to --- Published on [Forg](https://forg.to), a social network for people who create and ship on the internet. _HTML version: https://forg.to/articles/what-is-claude-mythos-and-why-it-will-not-be-available-to-public_