Security research team Calif announced they used Anthropic's unreleased Mythos Preview model to build the first public macOS kernel memory corruption exploit chain on an M5 Mac. MIE (Memory Integrity Enforcement) is Apple's flagship hardware security mechanism for M5 and A19 chips, five years and billions of dollars in the making. Its design goal was never absolute immunity but to make memory corruption exploits prohibitively expensive. Calif went from finding a flaw to a full exploit in five days.
The exploit chain uses two vulnerabilities and multiple techniques, starting from an unprivileged local user with only standard system calls and escalating to root. It is a pure data-driven kernel local privilege escalation, running on real hardware with macOS 26.4.1 and kernel MIE enabled.
Mythos excels at generalizing from one attack class to similar problems. It helped the team quickly locate defects belonging to known vulnerability classes, then human experts tackled the novel hardware defenses. This break-in validates the powerful combination of AI discovering bugs and humans bypassing defenses, proving that with top-tier large models, even a small security team can topple expensive technical barriers built by big companies.