AI HAR NETTOPP KNEKKT macOS PÅ 5 DAGER, Claude Mythos fra Anthropic knuser Apples M5 Security Fortress

While you were stacking sats and memeing about the next bull run, Anthropic quietly dropped Claude Mythos Preview, an ultra-restricted AI model available only to a tiny circle of partners through Project Glasswing. And this beast just proved it’s not playing around: it helped researchers build a working exploit that bypasses one of Apple’s most hardened security features on the latest M5 silicon.

Here’s what went down. In April 2026, the security team at Calif (a Palo Alto-based firm) was testing an early version of Mythos. Using the AI’s techniques, they discovered two brand-new, previously unknown bugs in macOS 26.4.1 running on Apple M5 chips. Then they chained those bugs together with a handful of clever memory tricks to create a data-only kernel local privilege escalation exploit.

Result? A normal user with zero admin rights can now corrupt kernel memory and get full root-level access to the entire system, bypassing Apple’s brand-new Memory Integrity Enforcement (MIE), a hardware-backed defense the company spent five years and billions of dollars building on top of ARM’s Memory Tagging Extension (MTE).

This wasn’t some theoretical proof-of-concept. Calif turned it into a working chain in just five days. They even drove the full 55-page technical report straight to Apple Park in Cupertino to hand-deliver it in person. PoC videos and the full breakdown are already public on their blog.

Apple is taking it seriously, they’re actively reviewing the report. A partial fix landed in macOS Tahoe 26.5 (released earlier this week), but the complete patch is still in the works. Calif notes that Mythos didn’t do it alone; human expertise was still required to finish the attack chain. But the speed is terrifying: what used to take months (or years) of manual reverse-engineering now happens in under a week with the right AI.

Why should crypto holders care?

Most of us run Macs for trading terminals, cold storage wallets, or MetaMask + hardware key setups. A kernel-level compromise means an attacker could silently grab seed phrases, clipboard data, USB-connected wallets, or even browser extensions holding your private keys, all without triggering alerts. In a world where AI is now finding thousands of high-severity zero-days across every major OS and browser, “assume breach” is no longer paranoia, it’s table stakes.

We’re officially stepping into Matrix/Terminator territory. AI isn’t just generating cute images anymore, it’s autonomously hunting zero-days that could own your entire rig.

Stay safe out there, update your macOS immediately, and maybe consider air-gapped cold storage for anything serious. The AI security arms race just went nuclear.