AI Model Chains Multi-Platform Flaws Into Working Exploits

Anthropic’s unreleased internal AI model identified and combined five separate vulnerability sequences across Chrome, Safari, Windows, and Linux into functional proof-of-concept exploits. On Chrome, four distinct flaws progressively disabled the JavaScript sandbox and achieved arbitrary code execution. Safari’s chain bypassed memory protections and disabled the same-origin policy without executing code, enabling credential and cookie theft. Windows exhibited denial-of-service conditions from kernel-level network handling. The Linux chain proved most severe: it leaked kernel addresses to bypass address space layout randomization, then executed root commands over the network by reusing a race condition.

Vendors patched all flaws under Anthropic’s 90-day coordinated disclosure timeline. Google and Apple formally credit Anthropic and Claude for specific fixes, providing some external validation, though the exact scope of AI autonomy versus human guidance remains undefined. The Windows and Linux findings rest primarily on Anthropic’s own documentation.

The work serves as both a capability demonstration and marketing for Anthropic’s Cyber Verification Program, which grants vetted security professionals access to Claude models. The company’s central claim—that AI can autonomously combine multiple vulnerabilities into functional exploits—has historically been the exclusive domain of specialized security teams.