Anthropic launches Glasswing and an AI model aimed at security

Anthropic has launched Project Glasswing, a security initiative aimed at defending critical infrastructure against AI-driven threats. At its centre is Claude Mythos Preview, a model specialised for cybersecurity, which in the company's internal testing autonomously identified and fixed a large number of high-severity software flaws across major operating systems and web browsers, including long-standing bugs that had evaded human detection. To limit the risk of such capabilities falling into the wrong hands, Anthropic has formed a defensive consortium of more than 40 industry participants, including Amazon, Google, Microsoft, Apple and CrowdStrike, to apply the model to both proprietary and open-source systems. Anthropic has committed 100 million dollars in model credits and four million dollars in direct donations to open-source security organisations.

What this means for your organisation

This points to vulnerability handling accelerating in both directions. Flaws will be found and fixed faster at the major vendors, but the same capability lowers the barrier for finding flaws with other intentions. For most organisations the consequence is practical: the window between a patch being released and it needing to be installed is shrinking. That is a governance challenge more than a technical one.

Berigo recommends

• Set a concrete deadline for how quickly critical updates must be deployed, and measure compliance against it.
• Decide deliberately how AI tooling will be used in your own security work, before someone decides it informally.
• Include AI-driven vulnerability discovery in the risk assessment for internet-facing services.
• Raise the topic with the board as a change of tempo, not as a single news item.

Source

• anthropic.com