In October 2026 Thompson’s always‑on Mac Mini was breached through CVE‑2026‑65400, a high‑severity flaw in the screen‑sharing service that lets attackers execute arbitrary code, gain root access and install a Monero crypto miner. He had enabled automatic security updates, but the fix arrived in a point release not covered by that setting, leaving him exposed for years under the mistaken belief that “install security updates automatically” would apply to all patches.
The image visually represents the compromise of Apple's macOS screen‑sharing service that led to unauthorized access and cryptocurrency mining on Ben Thompson's Mac Mini.
Thompson’s AI agent—a persistent Claude Code thread that logs his ideas and tracks projects—noticed that his account could suddenly run admin commands without a password, halted all operations and suggested defensive steps. Rather than disabling the agent, he used Claude to isolate the exact four‑second window of initial access, built a tool to watch for similar behavior, and wiped the compromised Mini, crediting the agent with giving him critical early warning that he might otherwise have missed.
The incident exposed a mismatch between macOS security architecture and agent workflows. Thompson criticized the Transparency, Consent, and Control (TCC) subsystem, which requires GUI‑based permission prompts invisible to headless agents, as operating at the wrong abstraction level and causing silent failures. He also noted Apple’s recent developer note tightening Full Disk Access controls, reflecting growing concern that broad agent access could endanger user data while agents need deep system access to function effectively.