San Francisco: Anthropic has unveiled Claude Mythos, its most powerful AI model to date, and announced it will not release it to the public. The reason: the model is too capable at finding and exploiting software vulnerabilities to be safely deployed at scale.
Over the past several weeks, Anthropic used a preview version of Mythos to scan critical software infrastructure. The model found thousands of zero-day vulnerabilities, previously unknown flaws, across every major operating system and every major web browser. Many of the bugs are decades old. Over 99% remain unpatched, which is why Anthropic cannot disclose details.
The capabilities were not intentional. Anthropic said it did not train Mythos for cybersecurity work. They emerged as a byproduct of general improvements in coding, reasoning, and autonomy. The same skills that make Mythos better at fixing code make it better at breaking it.
The scale of what Mythos can do is striking. Previous Claude models had a near-zero success rate at autonomous exploit development. Mythos converted 72.4% of known Firefox JavaScript vulnerabilities into working exploits. Engineers with no security training asked it to find remote code execution vulnerabilities overnight and woke to complete, working exploits.
In response, Anthropic launched Project Glasswing, a coordinated defensive effort to use Mythos to patch vulnerabilities before bad actors develop similar capabilities. Twelve core partners are involved, including AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorgan Chase, the Linux Foundation, Microsoft, and Nvidia. Forty organisations in total will receive access. Anthropic is backing the effort with USD 100 million in usage credits, along with USD 4 million in donations to open-source security organisations.
The announcement lands as Anthropic faces a complicated backdrop. The company has suffered two major security lapses in recent weeks, the accidental exposure of a draft Mythos blog post and the Claude Code source code leak via npm. It is also in a legal dispute with the US Department of Defense. Now it is asking the industry to trust it with one of the most dangerous AI capabilities ever disclosed.
Previous
