A U.S. security firm used Anthropic's Mitos AI model to bypass core macOS security mechanisms and gain administrator-level control [1, 2].

The breach demonstrates how generative AI can accelerate the discovery of system vulnerabilities, potentially rendering long-term security investments obsolete in a matter of days.

Researchers from the firm Calif conducted the study to test the resilience of Apple's operating system [1, 2]. The team said Thursday that they achieved the breakthrough within roughly five days of research effort [2]. This timeline stands in stark contrast to the development of the security architecture itself, which Apple spent five years building [2].

By utilizing the Mitos model, the researchers were able to elevate a normal-user process to a level of authority that allows full system control [1, 2]. This process involves identifying and exploiting flaws in the core security layers that were previously thought to be robust, a task that typically requires months of manual reverse engineering by human experts.

The findings suggest that AI can be used to automate the creation of exploits, allowing attackers to find paths through a system that human developers spent years securing [1, 2]. While the research was conducted by a security firm for demonstration purposes, the speed of the exploit raises questions about the future of software defense.

Apple has not yet provided a public response to the specific methodology used by the Calif research team [1, 2].

“The security was bypassed in five days”

This incident highlights a growing asymmetry in cybersecurity where AI-driven offensive capabilities can outpace traditional defensive development cycles. When a five-year security investment is compromised in five days, it suggests that static security architectures are increasingly vulnerable to automated analysis, necessitating a shift toward AI-powered real-time defense systems.