Anthropic's Mythos helped hackers crack Apple M5 defenses in five days

Apple spent five years and likely billions of dollars engineering a hardware-level shield designed to wipe out an entire category of memory attacks. A tiny research team with an AI just ripped through it in five business days.

Security firm Calif revealed how they built the first public kernel memory corruption exploit targeting Apple’s latest M5 silicon, leaning heavily on Anthropic’s specialized Mythos Preview cybersecurity model.

Breaking M5 in a week flat

The target here wasn't some legacy codebase. It was Apple’s brand-new Memory Integrity Enforcement (MIE), introduced last year across the iPhone 17 lineup and recently rolled out to Macs running M5 chips.

Built on top of Arm’s Memory Tagging Extension, MIE tags memory allocations with cryptographic secrets in hardware. If an app tries to access memory with the wrong tag, the chip halts the process immediately. Apple claimed MIE effectively neutralizes every modern iOS exploit kit floating around, including high-profile leaks like Coruna and Darksword.

Calif’s team broke it almost by accident. On April 25, researcher Bruce Dang spotted two bugs. Dion Blazakis joined Calif two days later, Josh Maine built out the tooling, and by May 1, they had a fully working local privilege escalation exploit.

Running against macOS 26.4.1 on bare-metal M5 hardware with kernel MIE fully enabled, the exploit starts from an unprivileged user shell, executes only normal system calls, and drops straight into root.

The model as a co-pilot, not an autopilot

The most revealing part of Calif’s post is how they actually leveraged Anthropic’s Mythos model.

According to Calif, Mythos is extraordinarily capable at generalizing across known bug classes. Once it understands the archetype of a vulnerability, it scans and flags instances almost instantly. That speed is how the team pinpointed the two underlying flaws in hours.

However, bypassing a cutting-edge mitigation like MIE wasn’t something the AI could do on autopilot. The human researchers had to design the exploit architecture and guide the bypass logic.

Even so, combining world-class offensive researchers with an AI trained specifically on exploit techniques shrank a multi-month engineering marathon into five days. The team drafted a 55-page technical breakdown, visited Apple Park to deliver the findings, and is holding the exploit details until Apple pushes a patch.

Why it matters

For everyday Mac users, there is no immediate danger. This requires local code execution, and Calif is following standard responsible disclosure protocols by waiting for Apple's upcoming security patch.

For defensive engineering, however, this is a seismic wake-up call. As Calif pointed out, Apple built MIE for a world that existed before Mythos.

Hardware mitigations take half a decade to architect, tape out, and ship at scale. If an advanced frontier model can help a three-person squad unravel that mitigation in less than a week, the traditional asymmetry favoring well-funded platform defenders has flipped.

We think this research cuts through the AI hype precisely because it didn't pretend the model did everything alone. It proved that pairing domain experts with targeted offensive models creates an unprecedented force multiplier against modern hardware mitigations.

The troubling takeaway is accessibility. If ethical researchers can achieve this in five days, well-resourced state actors and mercenary exploit brokers equipped with similar models will churn out zero-days faster than vendors can roll out silicon revs.