Calif breaks Apple M5 security with Claude Mythos

// 1h agoSECURITY INCIDENT

Calif breaks Apple M5 security with Claude Mythos

Security firm Calif developed a macOS kernel exploit that bypasses Apple's M5 Memory Integrity Enforcement. The local privilege escalation was chained in just five days using Anthropic's restricted Claude Mythos model.

// ANALYSIS

This is a watershed moment for AI-assisted offensive security, proving that frontier models can drastically compress the timeline for weaponizing zero-day vulnerabilities against hardened hardware.

–Calif bypassed Apple's flagship Memory Integrity Enforcement (MIE) on M5 silicon to achieve root access in macOS 26.4.1.
–Anthropic's Claude Mythos, accessed via the restricted Project Glasswing, assisted in discovering and chaining the two underlying kernel bugs.
–The five-day turnaround from initial bug discovery to a working exploit demonstrates the speed multiplier AI brings to vulnerability research.
–The incident underscores why models with advanced autonomous cyber capabilities remain tightly gated from public release.

// TAGS

califclaude-mythosapple-m5securityllmsafety

DISCOVERED

1h ago

2026-05-15

PUBLISHED

2h ago

2026-05-15

RELEVANCE

8/ 10

AUTHOR

Wes Roth

// KEEP READING

More AI developer news from the feed

EXPLORE FULL FEED

UPDATE10m ago

cmux hits v0.64 with session restore, subagents

The native macOS AI terminal releases v0.64, adding persistent session restore and native subagent panes for Codex Teams. A performance-focused alternative to Electron-based orchestrators, optimized for multi-agent workflows.

SECURITY2h ago

Claude Mythos uncovers macOS kernel exploit

Anthropic's unreleased Claude Mythos model assisted security researchers in discovering a critical privilege escalation vulnerability in macOS 26. The exploit bypasses Apple’s Memory Integrity Enforcement on M5 hardware, leading to a major security update for the Tahoe operating system.

BENCHMARK2h ago

Microsoft MDASH tops CyberGym with 100+ agents

Microsoft's Multi-model Agentic Scanning Harness (MDASH) coordinates over 100 specialized AI agents to automate the discovery and verification of software vulnerabilities. The system currently leads the UC Berkeley CyberGym benchmark with an 88.45% success rate, significantly outperforming Anthropic’s Mythos and OpenAI’s GPT-5.5 in generating functional proof-of-concept exploits.