Anthropic has announced its latest artificial intelligence model, called Claude Mythos Preview. The new model is the company’s most advanced one yet, surpassing its older Claude Opus 4.6 model as well as rival models such as Gemini 3.1 on various benchmarks. However, Claude Mythos Preview may be so powerful that Anthropic is not ready to release it yet, and the reason is cybersecurity.
The Dario Amodei-led claims that Claude Mythos is so powerful that it was able to spot thousands of cybersecurity vulnerabilities that were missed by humans across apps and operating systems.
As a result, Anthropic has announced that it is working with over 40 tech companies to use Mythos and fix such vulnerabilities before potentially releasing such a model to the public.
Claude Mythos found cybersecurity flaws humans couldn’t
The AI startup said in a blog post that Claude Mythos Preview could identify “thousands of zero-day vulnerabilities” across web browsers, and operating systems, which could pose a big risk in terms of cybersecurity. Zero-day vulnerabilities refer to flaws that were not identified by software developers previously.
This included a 27-year-old vulnerability in OpenBSD, an open-source operating system known for security-first approach, often used in firewalls and servers. Claude Mythos Preview also found numerous vulnerabilities in Linux kernel, which is used to run most servers worldwide. Anthropic claimed that these flaws could allow an attacker to “escalate from ordinary user access to complete control of the machine.”
Do note that Claude Mythos was first spotted in a data leak of Anthropic’s content management system (CMS). The company had confirmed at the time that it was going to be its most powerful AI model to date.
Anthropic launches Project Glasswing
Following the observations, Anthropic decided to delay, if not cancel, the release of Claude Mythos to the general public. Instead, the company has launched Project Glasswing, where it will work with major tech companies to identify and patch such vulnerabilities across various apps and operating systems.
Anthropic CEO Dario Amodei wrote on X, “The dangers of getting this wrong are obvious, but if we get it right, there is a real opportunity to create a fundamentally more secure internet and world than we had before the advent of AI-powered cyber capabilities.”
The consortium includes major companies such as Apple, Microsoft, Google, Amazon Web Services, CrowdStrike, Palo Alto Networks, Cisco, Broadcom, and the Linux Foundation.
Anthropic has committed up to $100 million in model usage credits to support Project Glasswing efforts, with partners paying for usage beyond this allocation.
Why is Project Glasswing important?
Anthropic claims that Claude Mythos is just one example of how advanced AI models have become. The company believes that it is necessary to fix such vulnerabilities right now, before such models become more widespread.
The company’s official X account wrote, “Given the pace of AI progress, it won’t be long before models this capable are widespread.”
Do note that Claude Mythos was not specifically trained for cybersecurity but benefits from its advanced coding and reasoning skills. According to Anthropic, the improved cyber capabilities are a byproduct of its general-purpose design.
Anthropic is collaborating with U.S. government agencies and security organisations to develop safeguards and best practices for deploying such powerful models responsibly.
The Dario Amodei-led startup had a fallout with the US Pentagon after it walked away from a defense deal over disagreements on unrestricted AI use earlier this year.
