Cybersecurity researchers from Calif have successfully bypassed Apple’s Memory Integrity Enforcement (MIE) mechanism in macOS using the well-known AI model Claude Mythos, as detailed in a recent report.
The team discovered two vulnerabilities in macOS in April and linked them to create an attack chain that allowed a regular local user to gain root access—maximum privileges within the system.
The breach was tested on a real device equipped with an M5 chip, where MIE protection was enabled.
Experts have informed Apple about the issue and plan to publish technical details once it has been resolved.
They believe that Mythos could not have executed the attack alone without human assistance. While it can replicate known methods, it lacks the ability to devise new ones.
Apple Security
Memory-related errors remain the most common type of vulnerabilities across various platforms, including iOS and macOS. A fundamental principle in cybersecurity is that if a problem cannot be completely prevented, one must accept the risk and work to mitigate it. One approach is to make attacks significantly more complex and costly for the attacker.
Implementing such protective measures can be expensive. Calif emphasized that Apple operates intelligently, controlling the entire system—from hardware to software. Consequently, the company has integrated many security mechanisms directly into its chips, making them much harder to bypass. Many security experts consider Apple devices to be the most secure mass-market platform for everyday users.
The latest example is MIE, a memory protection system that leverages the hardware capabilities of the chip. It was introduced as a key security feature in the Apple M5 and A19 chips, specifically designed to thwart attacks based on memory errors.
Apple spent five years developing the system. MIE disrupts all publicly known attack chains against modern versions of iOS, including recently leaked toolkits like Coruna and Darksword.
Calif explored how AI can assist in creating attacks that remain effective even with such memory protection in place.
The method to hack macOS was discovered by chance. Bruce Dang identified the vulnerabilities on April 25, and Josh Main developed the necessary tools, resulting in a working version by May 1.
“This experiment illustrates what we might face in the future. Apple developed MIE in a world where Mythos Preview did not yet exist. Now we will see how the world’s best protection technology withstands the first major wave of vulnerabilities discovered with AI,” the experts concluded.
It’s worth noting that in April, Anthropic withdrew Mythos from public access due to significant security risks.