When AI Agents Go Wild: The Hugging Face Breach #DTF054

Can we actually control autonomous AI agents once they leave the sandbox?

In Episode 54 of the DTF Cyber Podcast, Damian, Troy, and Fern dive into real-world instances of AI models breaking containment and acting beyond human control.

From an AI benchmark escaping its isolated test environment to laterally move and hack external infrastructure like Hugging Face, to the viral whistleblower tweet from former OpenAI and Anthropic safety researcher Jacob Coxon warning that labs are "racing straight to self-improving superintelligence", the crew breaks down what this shift means for real-world enterprise cybersecurity.

Timestamps:

00:00 – An AI model breaks out of its sandbox

00:36 – Intro: When the machines go wild, are we losing control?

00:54 – "AI Gone Wild" & retro tech memories

02:18 – The Hugging Face Incident: Breaking containment at machine speed

04:16 – Why traditional allowlists and denylists fail against autonomous agents

06:00 – Tech leadership culture: Do Dario and Sam prioritize security?

08:14 – The nation-state race: Can we afford to slow down?

11:30 – Media hype, FUD, and the battle over AI data centers & power