Anthropic Internal Model Chained Four Chrome Flaws Into macOS Sandbox Escape With Minimal Prompting

The model produced five exploit chains after initial prompting with little documented human input. In Chrome it chained V8 memory corruption, sandbox escape, CacheStorage race condition, and FedCM use-after-free to achieve arbitrary code execution. Safari chain disabled same-origin policy via memory read-write. Android extracted data from locked devices over USB. Windows delivered HTTP/3 denial-of-service. Linux achieved root via multipath TCP. Anthropic claims 62 critical or high CVEs assigned to Google, 13 sandbox escapes among them. Vendor release notes credit specific researchers assisted by Claude but provide no independent audit trail of model autonomy versus human steering at each step. The Windows CVE record lists no finder, breaking the attribution chain. This pattern indicates AI can compress the timeline from flaw discovery to chained proof-of-concept that previously required small specialist teams, yet the absence of granular logs leaves open whether the model performed end-to-end synthesis or merely accelerated isolated tasks. Responsible disclosure occurred, but the capability demonstration itself lowers the barrier for future malicious actors once similar models proliferate beyond lab controls. Vendors now face incoming AI-generated reports at higher volume; triage systems must adapt to distinguish model-generated noise from high-fidelity chains. Expect increased investment in automated verification of AI-submitted reports and tighter controls on internal model weights to prevent leakage of the chaining methodology.