Contexte

Anthropic has introduced Project Glasswing, a restricted defensive initiative centered on its most powerful model to date, Claude Mythos Preview. This frontier AI possesses unprecedented capabilities in cybersécurité, including the autonomous discovery of thousands of critical “zero-day” vulnerabilities across major operating systems and browsers. Due to its potential for offensive misuse, Anthropic has rejected a public release, instead granting controlled access to a consortium of twelve partners such as AWS, Google, and Microsoft. The program includes $100 million in credits and direct donations to bolster open-source security infrastructure and mission-critical codebases. While the model excels at defensive reasoning, its development has sparked legal and political friction with the U.S. Department of Defense over lethal military applications. Ultimately, Project Glasswing aims to accelerate the patching of global digital infrastructure while establishing a cautious new standard for the deployment of highly capable AI.

Sources