Anthropic has developed 'Claude Mythos Preview,' an AI with exceptionally high cyberattack capabilities, and has also launched 'Project Glasswing,' which will provide a preview version to companies such as Microsoft and Apple.
On April 7, 2026, Anthropic announced the release of ' Clude Mythos Preview ' to select organizations. Claude Mythos Preview boasts an extremely high capability to 'discover and exploit software vulnerabilities,' and has already uncovered a large number of critical vulnerabilities in Linux and OpenBSD. In response to the increasing security risks associated with the advancement of AI, Anthropic has also launched ' Project Glasswing ,' a project supporting organizations such as Microsoft, Apple, and the Linux Foundation.
Claude Mythos Preview\red.anthropic.com
Project Glasswing: Securing critical software for the AI era \ Anthropic
https://www.anthropic.com/glasswing
The following diagram compares the performance of Claude Mythos Preview and Claude Opus 4.6 as coding agents. Claude Mythos Preview shows a significant performance improvement compared to Claude Opus 4.6.
Inference performance has also improved significantly.
The PC's automated operation capabilities have also improved.
Anthropic highlights Claude Mythos Preview's remarkable ability to discover and exploit software vulnerabilities. According to Anthropic, Claude Mythos Preview's vulnerability detection capabilities are 'at a level that surpasses most humans, with the exception of a handful.'
The following are the results of running the task 'Create an exploit by exploiting a Firefox vulnerability' with 'Claude Sonnet 4.6', 'Claude Opus 4.6', and 'Claude Mythos Preview'. While the majority of Claude Sonnet 4.6 and Claude Opus 4.6 show gray bars indicating failure, Claude Mythos Preview created executable exploits at a high rate of 72.4%.
Furthermore, Claude Mythos Preview has successfully discovered thousands of vulnerabilities in real-world software and autonomously created exploits without human intervention. Examples of its achievements are as follows:
A vulnerability has been discovered in OpenBSD that has been lurking for 27 years. An attacker can cause it to crash simply by connecting to any machine.
A vulnerability that had been lurking in FFmpeg for 16 years was discovered. This vulnerability had remained undetected even after more than 5 million scans using automated testing tools.
- Discovered multiple vulnerabilities in the Linux kernel and created an exploit that combines them to gain full access privileges.
Anthropic pointed out that 'given the speed of AI development, it's only a matter of time before AI with a high risk of misuse is widely released,' and launched 'Project Glasswing' as an urgent effort to use AI for defensive purposes before attackers begin to exploit it.
Project Glasswing is an initiative that grants multiple organizations access to AI models such as Claude Mythos Preview to help them discover vulnerabilities. Participating organizations include AWS, Anthropic, Apple, Broadcom, Cisco, CrowdStrike, Google, JP Morgan Chase, Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks, and is expected to expand to more than 40 organizations in the future.
Project Glasswing will provide participating organizations with AI usage credits worth up to $100 million (approximately 15.9 billion yen). The fee for participating organizations to use the Claude Mythos Preview is $25 (approximately 3,960 yen) for input and $125 (approximately 19,800 yen) for output per million tokens. It has also been announced that $2.5 million (approximately 396 million yen) will be provided to Alpha-Omega and OpenSSF, and $1.5 million (approximately 238 million yen) to the Apache Software Foundation.
We're committing up to $100M in Mythos Preview usage credits for our partners and over 40 other organizations that maintain critical software, including open-source projects.
— Anthropic (@AnthropicAI) April 7, 2026
Anthropic will report back what we learn.
Anthropic is reportedly in discussions with government agencies regarding the cyberattack and defense capabilities of Claude Mythos Preview. There are no plans to release Claude Mythos Preview to the public, and safeguards are being developed to prevent dangerous output.
We do not plan to make Mythos Preview generally available. Our goal is to deploy Mythos-class models safely at scale, but first we need safeguards that reliably block their most dangerous outputs.
— Anthropic (@AnthropicAI) April 7, 2026
We'll begin testing those safeguards with an upcoming Claude Opus model.
Additionally, a system card containing detailed information about Claude Mythos Preview is available at the following link.
Claude Mythos Preview System Card
(PDF file) https://www-cdn.anthropic.com/8b8380204f74670be75e81c820ca8dda846ab289.pdf
Related Posts:
