Anthropic releases Mythos AI model; Google, Apple and Microsoft join the effort

Anthropic, the owner of OpenClaw, is taking steps to equip some of the world’s largest technology groups with advanced tools to detect and remediate vulnerabilities in both hardware and software.

Accordingly, the company has provided a beta of its new AI model named Mythos to about 50 corporations and organizations operating infrastructure worldwide. The list includes giants such as Amazon, Microsoft, Apple, Google (Alphabet), and the Linux Foundation.

Why Anthropic is rolling out Mythos

The move comes as cybersecurity researchers and software developers raise concerns about AI’s ability to exploit security vulnerabilities. Many experts warn that AI models are advancing rapidly and could enable large-scale disruptions to online operations if misused.

At the same time, security experts say AI models can uncover software defects at speed, and Anthropic’s effort is expected to help companies stay “one step ahead” of cybercrime and other threats.

What Mythos is designed to do

Security experts say AI can detect a large number of software faults in a short time, which can speed up patching and strengthen systems.

Mr. Logan Graham, head of Anthropic’s Frontier Red Team, the risk assessment unit for models like Claude, said Mythos has demonstrated the capability to perform dangerous tasks, including searching for and exploiting software vulnerabilities. However, he said the company currently has no plans to publish the model publicly.

“We need to determine whether it can be deployed safely, and there is no definitive answer yet,” Graham emphasized.

Rising concerns: faster detection to exploitation

In the past six months, concerns have intensified alongside the rapid development of AI applications. Researchers say AI not only improves at detecting flaws, but also shortens the time from detection to exploitation.

Earlier research from Stanford University found that AI has nearly reached human levels in detecting and exploiting vulnerabilities on real-world systems. The article also cites an example earlier this year: Anthropic’s Claude Opus 4.6 model found a number of critical flaws on the Firefox browser within two weeks, exceeding the number of bugs the global community typically reports in two months.

Performance and cost comparisons

According to Mr. Graham, Mythos is about 10x more effective than earlier AI models in finding vulnerabilities, in terms of cost parity.

Industry impact and next steps

Although there are no commercialization plans yet, Anthropic believes other models could achieve similar capabilities within a few years. That outlook creates an urgent demand for the tech industry to prepare for a future in which the gap between detection and exploitation of vulnerabilities is nearly zero.

“We need to start today to adapt to that world,” Graham warned.

According to WSJ