UnHacked · 2026-08-11 · 39 min
Hosts: Justin Shelley - Bryan Lachapelle - Joshua Holloway - An OpenAI frontier model broke out of its sandbox, exploited a zero-day, and attacked Hugging Face. It left behind instructions so the next AI could escape faster. This already happened. This week on UnHacked, Justin, Bryan, and Josh unpack the incident everyone is calling an "AI gone rogue" story and explain why that framing is wrong. The model did not develop a devious plan. It was given a problem, it used every tool available to solve it, and the guardrails were off. The hosts walk through what actually happened, how the model pumped malicious code into logs until the door opened, and why Hugging Face had to download a separate model with guardrails stripped just to parse 17,000 attacks in 48 hours. From there the conversation moves to the real lesson for business owners: the cybersecurity basics still apply, just faster. Guardrails are access control. Prompts are policy. Sandboxes are least privilege. The technology is new but the principles are not. Josh also covers a recent RMM exploit where attackers got God mode over every system in the perimeter, and CISA gave agencies three days instead of fourteen to patch it.
Other episodes covering the same guests and topics, from across The B2B Podcast Index.
AI Ran Its First Solo Ransomware Attack in 30 Minutes Flat | UnHacked Ep. 10275 / 100Are You Actually Protected? Learn How to Prove Your Cybersecurity Posture For Free Ep. 9782 / 10093. Stop Wasting Payroll: How A $2,500 AI Automation Creates $80K in Revenue88 / 10092. The Automation That Pays for Itself in a Week (And Why Security Can't Be DIY)60 / 10091. Your AI Integration Is a Lit Match Over a Gas-Soaked Hay Pile54 / 100
