AI Hacker Shock

Anthropic has been running limited testing of the unreleased Claude Mythos Preview inside Project Glasswing, highlighting that AI can find vulnerabilities at scale that humans have previously missed.

According to the core claim, Mythos found thousands of zero-days within a few weeks, and reached the level of stitching multiple vulnerabilities together to build an exploit chain from a single prompt.

Citing this level of danger, Anthropic has not released the model publicly, and instead has given limited access to about 40 organizations, including Apple, Google, Microsoft, and the Linux Foundation, for use in defense and patching.

The piece argues this result could actually be a reversal for the tech talent market.

• While the trend of AI replacing coding work has continued, it has also revealed just how large the security debt in existing software is.
• As a result, demand for roles such as cloud security, AI red-teaming, vulnerability management, and security architecture could grow.
• Citing Wiz, the piece argues there are 12-18 months before similarly capable models spread widely as open source, and that defense teams need to be built up during that window.

However, outside experts recommend caution regarding Anthropic's unverifiable claims. Since it is not a public model, independent reproduction is difficult, and currently the claims rely on reports from a limited group of partners. Even so, the piece strongly pushes the paradox that "AI may reduce the number of developers while simultaneously creating even greater demand for security personnel."