Anthropic's Mythos Model Breaks Apple's M5 Memory Defense in Five Days

AIMPACT message, May 16 (UTC+8): According to monitoring by Beating, the security research team Calif announced that they successfully constructed the first publicly disclosed macOS kernel memory corruption exploit chain on M5-chip Mac devices, leveraging Anthropic’s unreleased Mythos Preview model. MIE (Memory Integrity Enforcement) is Apple’s flagship hardware security mechanism designed for M5 and A19 chips, involving five years of development and billions of dollars in investment. Its purpose is not absolute immunity from hackers, but rather to mitigate memory corruption vulnerabilities by significantly increasing exploitation costs. Yet, the Calif team achieved full exploitation in just five days from vulnerability discovery. This attack chain combines two vulnerabilities and multiple techniques, starting from an unprivileged local user account and relying solely on standard system calls to ultimately gain root privileges on the device. The exploit is a purely data-driven kernel local privilege escalation targeting real, bare-metal hardware running macOS 26.4.1 with kernel MIE enabled. The Mythos model excels at rapidly generalizing from one type of attack to similar problems, helping the team quickly identify vulnerabilities belonging to known classes, after which human experts overcame the novel hardware defenses. This breach validates the efficiency of the “AI discovers vulnerabilities + experts bypass defenses” approach, demonstrating that with the assistance of top-tier large models, small security teams can challenge the multi-billion-dollar technological barriers built by major corporations. (Source: BlockBeats)