An elevation-of-privilege exploit has emerged that bypasses Apple’s next-generation memory protection technology, MIE (Memory Integrity Enforcement), developed over five years. The attack was developed in a short period using an artificial intelligence model, raising concerns that an era of AI-driven cyberattacks is becoming a reality.

On May 20 (local time), Japanese outlet ITmedia reported that U.S. security company CalypsoAI said it used a preview version of Anthropic’s AI model Mithos to complete attack code targeting M5 chip-based Macs in just five days.

The exploit targeted M5 chip-based Macs running macOS 26.4.1 (25E253) in a bare-metal environment with MIE enabled. CalypsoAI said it combined two macOS vulnerabilities and several attack techniques, starting from an unprivileged local user account and ultimately gaining root privileges using only system calls.

The vulnerabilities were found on April 25, and exploit development was completed on May 1. CalypsoAI said it has already provided technical details to Apple headquarters and plans to release a 55-page technical report and a proof-of-concept video after Apple distributes a security patch.

MIE is a memory protection technology Apple unveiled in September last year. It was developed over about five years based on Arm’s Memory Tagging Extension, and Apple introduced it at the time as an “unprecedented concentration of design and engineering”. The core is to combine Apple silicon hardware and operating system security to implement an always-on memory safety protection feature across a range of devices for the first time in the industry.

Apple products have been seen as a difficult platform to attack thanks to a structure that integrates security functions from the hardware stage. MIE in particular has been cited as a key defense that blocks exploitation of memory corruption flaws. CalypsoAI claimed, “To the best of our knowledge, this is the first disclosed case of an exploit targeting a macOS kernel on hardware equipped with MIE defenses.”

The focus of the research was on how deeply AI can be involved in vulnerability discovery and exploit development. CalypsoAI said Mithos learned specific types of attack methods and could apply them across similar classes of vulnerabilities. It added that the reason it found the vulnerabilities quickly was also because they were from a previously known class of issues.

CalypsoAI drew a line, saying the attack was not completed by AI alone. “Because MIE is a new state-of-the-art defense technology, AI had limits in autonomously bypassing it, and human expert knowledge was needed in the final stage,” it said.

CalypsoAI is a company founded by former Google security researchers, and it is researching AI-based attack and defense technologies with Anthropic and OpenAI, among others.

In the industry, analysts say the case could be a signal flare that competition in AI-driven cyberattacks is beginning in earnest. As AI accelerates vulnerability discovery and attack automation, even cutting-edge hardware security technologies are being put to the test.