Sandboxes Are Optional Now, Apparently

Late July brought a wave of frontier AI agents slipping their (very loose) leashes: an OpenAI model wandered loose inside Hugging Face's infrastructure for a week, Anthropic and Meta admitted to similar incidents, and UK AISI clocked their own model breaking containment during an independent eval. Meanwhile, a home user's OpenClaw agent quietly hijacked a stranger's gym booking. We dig into why "the whole internet is held together by safety pins," whether billion-dollar labs get to plead "security is hard," why software engineering mostly isn't real engineering, and what any of this means for the coming fight over open-weight models.

Chapters

(00:00) - Intro

(01:25) - The OpenAI/Hugging Face Containment Escape

(09:54) - Cost Centers vs. Profit Centers (Praise for AISI)

(17:37) - "You Do Not Get to Fuck Up This Much"

(22:17) - Doomers, Utter Alignment Failure & How Hacks Actually Happen

(37:53) - "Forcing the End" and What Can Anyone Actually Do?

(47:34) - Igor's Pivot to Open-Weight Models and Steelmanning the Case Against Them

(58:01) - Outro & Sign-off

Critical Links

Below are the most important links for this episode. For more, visit the episode page on Kairos.fm.Hugging Face blog - Anatomy of a Frontier Lab Agent Intrusion: A Technical Timeline of the July 2026 IncidentMETR report - Brief independent investigation of agents' behavior, reasoning and collaboration in the OpenAI / Hugging Face hacking incidentUK AI Security Institute incident report - Incident Report: unsanctioned agent behaviour during cyber testingHillel Wayne essay - Are We Really Engineers?Deutsche Welle article - OpenAI apologizes for not reporting Canada mass shooterPBS Frontline essay - Forcing the EndFelonyBench websiteTaylor Troesh (satirical) blogpost - Leaving OpenAI