Anthropic is limiting initial access to its new Claude Mythos Preview model to a curated consortium of companies. The decision follows internal testing where the model identified thousands of critical, often decades-old, security flaws across major operating systems, web browsers, and core libraries like OpenBSD, FreeBSD, and the Linux kernel. The company stated that 99% of these vulnerabilities remain unpatched, making broader release an unacceptable risk.

The technical architecture of the response is Project Glasswing, an alliance with over 40 firms including AWS, Apple, and Google. The engineering trade-off is clear: restrict a general-purpose model's availability to use its capabilities defensively at scale. The system will automate the discovery and patching of bugs—a process historically dependent on scarce expert labor—before malicious actors can exploit them.

This signals a transitional phase where offensive and defensive AI capabilities are in a race. While Anthropic predicts AI will ultimately lead to more hardened software, the immediate path involves managed deployment. The model's findings, which include weaknesses in cryptography protocols like TLS and AES-GCM, suggest that the attack surface for both legacy and modern systems is far larger than previously understood.

Source: CoinTelegraph