Anthropic has launched Project Glasswing, a cybersecurity initiative that is making its AI model Claude Mythos Preview, which specialises in detecting software security vulnerabilities, available to selected companies, CNBC reported on April 7.

Apple, Google, Microsoft, Nvidia and Amazon Web Services (AWS) are participating as initial partners and will use the model for defensive security work. More than 40 companies, including CrowdStrike and Palo Alto Networks, are also taking part.

Diane Penn, head of research product management at Anthropic, said the project was "a first step to give cyber defenders a chance to respond proactively in an area that will become increasingly important". Chief Executive Dario Amodei (Dario Amodei) wrote on X (Twitter) that "the risks are clear if it goes wrong, but if implemented correctly, it can create a fundamentally safer internet and world than before AI".

Claude Mythos Preview is not a model trained exclusively for cybersecurity but a general-purpose model that includes coding and reasoning capabilities. Anthropic said the Mythos model found bugs that had been difficult to detect previously, including a 27-year-old OpenBSD bug.

Anthropic plans for now not to release the model to the public. It aims to use the limited rollout to test ways to deploy Mythos-level models at scale while blocking the risk of misuse.

All companies in Project Glasswing build or maintain core software infrastructure and will use the model to strengthen security for their own systems and open-source systems. Anthropic will provide up to $100 million in usage credits for this.

Newton Cheng, head of cyber at Anthropic Frontier Red Team, said the goal was "to help companies get familiar with this capability before they can use it broadly".