Anthropic Restricts Powerful New AI Model, Citing Unprecedented Security Risks

Anthropic has made a striking decision: its newly developed AI model, Claude Mythos Preview, will be released to a limited set of partners. The reason is as impressive as it is alarming. During testing, the model identified thousands of critical security flaws across every major operating system and web browser, suggesting a capability that could be dangerously misused.

The company pointed to a sharp rise in AI-powered cyberattacks and argued that broad release of such a potent tool would be reckless. "Given the rate of AI progress, it will not be long before such capabilities proliferate," an Anthropic statement read, highlighting the risk of these powers falling into the wrong hands.

In response, Anthropic launched Project Glasswing, a coalition of over 40 major firms including Amazon, Apple, Google, Microsoft, and Nvidia. The initiative will harness the Claude Mythos Preview model defensively. The goal is to systematically find software bugs, share intelligence within the alliance, and patch vulnerabilities before they can be weaponized.

The scale of the discovery is notable. The AI uncovered flaws that are often subtle and have evaded detection for years—some for over a decade. It found a now-patched 27-year-old bug in the security-focused OpenBSD system, a 16-year-old issue in the FFmpeg library, and numerous weaknesses in the Linux kernel and common cryptography protocols.

Anthropic stated that 99% of the vulnerabilities it located remain unpatched, preventing public disclosure. The company acknowledges a difficult transition period ahead, where offensive uses of AI may surge. However, its long-term bet is that AI-driven defense will ultimately prevail, leading to fundamentally more secure software. The work, they concede, will take years.

Source: CoinTelegraph