“Hanlon's razor is just [OpenAI] was given so much money they don't have to give a shit. As long as they have FU money they will continue to behave in this way.”

Mike and Nick return to OpenAI's hack of Hugging Face to explain why agents weren't 'going rogue.' The episode concludes with the national security perspective.

• The recap: AI agents that hacked Hugging Face were trained to hack and collaborate, then handed impossible tasks in an insecure environment. “Breakout” was inevitable; this was negligence and user error, not intelligence.
• Against Instrumental Convergence/going rouge: The doomer framing collapses model, harness, and environment into a single agent with intent. Mike argues data science needs no drives: models are stateless composites steered by context, which is why swapping the harness or system prompt changes behavior.
• The Adults in the Room: Nick argues qualified cybersecurity and national security experts exist but are sidelined by executive politics, while Chinese open source models shape the global landscape. Europe's open source ecosystem and secure enclaves may be a way forward.

Timestamps

0:00 Intro

1:12 "going rouge" and "instrumental convergence"

12:11 Why doesn't the public understand this?

30:31 METR/Redwood investigation

35:14 AI models as composites

43:46 Why didn't the cyber professionals show up?

48:56 What is cybersecurity?

56:05 Could cyber investigations end the bubble?

01:04:24 Adults in the room and open source

01:16:50 Silicon Valley's power and the future

Deeper learning

"Grandpa Shelby" post: https://misaligned.markets/understanding-large-language-models/

Mike's reporting on the incident: https://misaligned.markets/antidote-to-hype-rogue-ai-agents/

Computer scientist Cal Newport on "going rouge:" https://calnewport.com/has-ai-gone-rogue/

Lee Vinsel on Criti-hype: https://freedium-mirror.cfd/https://sts-news.medium.com/youre-doing-it-wrong-notes-on-criticism-and-technology-hype-18b08b4307e5

Eryk Salvaggio on stocastic flocks: https://mail.cyberneticforests.com/models-dont-go-rogue/

Zitron on OpenAI's possible death: https://www.wheresyoured.at/what-happens-if-openai-dies/

OpenAI execs fleeing: https://www.businessinsider.com/executives-who-left-openai-in-2026-8

Reach out

Youtube: https://www.youtube.com/@LastEnclosure

Bluesky: bsky.app/profile/lastenclosure.bsky.social

Site: rss.com/podcasts/the-last-enclosure

“This is what they're doing with their billions of dollars. They're burning compute to get a model to talk on a fake message board to hack other companies and going: huh, neat.”

Mike and Nick analyze OpenAI's July incident where training agents breached Hugging Face, arguing it was a result of negligence rather than emerging AI autonomy. They explore how reinforcement learning, environmental attractors, and stigmergy created the appearance of coordination. The episode warns against 'doomer' narratives that serve corporate PR and obscures the need for practical regulation.

• The 'Rogue AI' Framing: OpenAI presented the Hugging Face breach as a sign of impressive emergent capabilities, but this is PR meant to hype AI's power and evade accountability.
• Stigmergy and Attractors: The agents didn't intentionally collaborate they converged on Artifactory as an 'attractor state' and left 'breadcrumbs' that shaped future agent beahvior.
• Systemic Negligence: OpenAI failed at basic cybersecurity by using a shared Artifactory instance across agent boxes and lacking observability for two months.

Timestamps

0:04 Introduction to The Last Enclosure

3:35 The OpenAI/Hugging Face incident overview

8:15 ExploitGym benchmark and the Artifactory exploit

15:19 Subagents and the accidental message board

22:09 Timeline of the hack and OpenAI's failure to contain

31:27 Analyzing the 'emergent behavior' narrative

35:06 Critique of the AI bubble and marginal utility

48:48 Stigmergy and the ant mound metaphor

58:29 Coordination games and Schelling points

1:06:31 The 'Four Paradoxes of Capitalism' and market attractors

1:13:23 Enumerating OpenAI's negligence

1:27:56 Closing and upcoming national security episode

Deeper learning

Mike's reporting on the incident: https://misaligned.markets/antidote-to-hype-rogue-ai-agents/

Timeline of OpenAI hack: https://ericboyd.com/articles/openai-hugging-face-incident-black-hat-2026#the-accidental-agent-message-board

Technical breakdown: https://cyberwarrior76.substack.com/p/the-openai-hugging-face-exploitgym

Other “rogue” AI incidents: https://www.cnbc.com/2026/08/09/israeli-startup-irregular-linked-to-ai-hacks-openai-anthropic-meta.html

Video illustration of how reinforcement learning works: https://www.youtube.com/watch?v=L_4BPjLBF4E

NIST cybersecurity glossary: https://csrc.nist.gov/glossary

Reach out

Youtube: https://www.youtube.com/@LastEnclosure

Bluesky: bsky.app/profile/lastenclosure.bsky.social

Site: rss.com/podcasts/the-last-enclosure

"AI is like the guy that showed up and started pissing in everyone's beer at the party."

Mike categorizes various groups resisting or critically engaging with generative AI, from total abstainers to 'scopers' who seek responsible use cases.

• Global harms: While AI has become a hot button issue in the US, frustration in the states is a small part of a bigger movement.
• Diverse coalitions: Like the buckets of AI overlords from episode three, the resistance consists of many desperate groups banding together.
• Scoping AI: Mike briefly speaks to use cases that are narrow in scope and lower in their externalities to users and the public.

Deeper learning

The AI Resist List: airesistlist.org

Mike's thread on GenAI harms: bsky.app/profile/misaligned.markets/post/3migfnyvuak26

Ghost in the Machine Documentary: notaidoc.com

Ethical European model Apertus: infoworld.com/article/4051779/swiss-launch-open-source-ai-model-as-ethical-alternative-to-big-us-llms.html

EVO RNA GPT: hazyresearch.stanford.edu/blog/2024-03-14-evo

Inside Memphis’ fight against xAI: selc.org/news/inside-memphis-fight-against-xai/

Data centers in drought-region: newsweek.com/map-data-centers-built-drought-hit-areas-11997520

Jeff Geering AI and open source: jeffgeerling.com/blog/2026/ai-is-destroying-open-source

XKCD infrastructure meme: explainxkcd.com/wiki/index.php/2347:_Dependency

Distributed AI Research Institute: dair-institute.org

Dr. Fatima how to (Anti) AI better: youtube.com/watch?v=y85nqc2zm7M

Innocence Project LLM data science: hrdag.org/tech-notes/large-language-models-IPNO.html

Joy Buolamwini and the Algorithmic Justice League: https://www.ajl.org/about

Reach out

Site: rss.com/podcasts/the-last-enclosure

Bluesky: bsky.app/profile/lastenclosure.bsky.social