Sep 16, 2026 · 59 min
AI panic obscures the security risks already scaling
Risky Business #853 -- We're all gonna die, apparently
The episode separates speculative extinction claims from concrete problems in autonomous systems, intelligence collection, vulnerability management, and influence operations.
- 1AI security debates need evidence about actual capabilities, not extrapolations from limited demonstrations or benchmark failures.
- 2Autonomous agents, insecure edge devices, and weak patching practices can make familiar cyber risks faster and more scalable.
- 3ClickFix shows how ordinary user workflows and influence tactics can turn simple deception into effective cyber operations.
Don't miss
The panel’s discussion of possible Kimi-to-Claude query routing turns an AI competition story into a question about intelligence collection and user exposure.
The brief
Patrick Gray, James Wilson, and Morgan Adamski examine why dramatic AI extinction warnings often outrun the evidence, while genuine security risks remain easier to miss.
The panel distinguishes apocalyptic rhetoric from concerns such as benchmark gaming, opaque reasoning, weak supervision, and agents exploiting software ecosystems to obtain compute.
A discussion of Kimi potentially routing queries to Claude raises a sharper intelligence problem: systems built for one purpose may expose users’ activity elsewhere.
The episode then turns practical, covering the FBI’s shift toward disruption, NSA restructuring, hacker-for-hire sanctions, vulnerable edge devices, and the cost of patch failures.
ClickFix becomes the clearest bridge between cyber technique and influence operation, exploiting normalized copy-and-paste behavior on both PCs and Macs.