Agents want deep system access and the security bill is arriving
Greg Kroah-Hartman's Kernel Recipes talk on security in the LLM age drew praise for its candor. Commenters pulled out the numbers on Mythos finding dozens of vulnerabilities, and one noted it all boiled down to about an hour of kernel development work. At the same time, Apple is changing Full Disk Access in macOS, and the thread revolves around Meta's Muse, which sends users to the Full Disk Access pane to grant itself broad reach. Reactions ranged from 'macOS is already almost unusable due to permission prompts' to excitement about Muse Gadgets connecting to a real agent, tempered by 'this being Facebook is hard to trust'.
The pattern: AI is making vulnerability discovery cheap on the attacker side, while consumer agents are asking for the keys to the whole machine. Operating system vendors respond with more prompts.
The counterpoint is that Kroah-Hartman's data is checkable because it is the Linux kernel, a point commenters made about trusting it over closed-vendor claims.
So what?
If you ship an agent that asks for broad permissions, expect operating system vendors to add friction and users to hesitate. Narrow scopes and clear explanations of why you need access will convert better. If you maintain open source or run infrastructure, assume faster vulnerability discovery and plan your patching around it.