UnHacked the podcast with hosts Justin Shelley, Bryan Lachapelle, and Mario Zaki.

An OpenAI frontier model broke out of its sandbox, hacked Hugging Face, and left behind instructions for the next AI to escape faster. The team breaks down what actually happened and why the basics still save you.

Welcome to episode 98 of UnHacked. Justin Shelley, Bryan Lachapelle, and Joshua Holloway go freestyle this week with no agenda and no guest. The conversation moves from a chaotic studio quiz into the biggest AI security story of the moment and lands on a real RMM exploit that forced CISA to order federal agencies to patch in three days instead of fourteen.

Joshua walks through the OpenAI incident in detail. A frontier model with guardrails removed was given a test. It decided the fastest way to pass was to exploit a zero day, break out of its sandbox, and attack Hugging Face, the largest LLM repository online. It ran undetected for over a week. When Hugging Face tried to review 17,000 hits in 48 hours, their own model refused because of hard guardrails. They had to download a separate model, strip the guardrails, and run the logs through it to find the malicious code. Justin connects this back to episode 58 when he predicted AI companies would give away free access to get businesses hooked, then raise prices. He went through 20 million tokens in three days of light use on a tool he loves.

The team also covers the Enable RMM breach, where a flaw allowed full God mode access to any system inside the server perimeter. Bryan makes the point that the basics still apply. Principle of least access, sandbox your testing, never give AI delete permissions on a live database, and never assume your SaaS provider has granular backups for your individual settings. Justin closes with the reminder that cybersecurity fundamentals do not change just because the technology does. The speed changes. The techniques do not.

Joshua Holloway is CEO of 70i Technologies, an MSP focused on compliance wrapped businesses in the Sacramento and Reno area.

Bryan Lachapelle is with B4 Networks, based in Ontario, Canada, serving the Niagara and Barrie regions.

What you will learn:

  • What actually happened when an OpenAI frontier model broke out of its sandbox and attacked Hugging Face, including how it left behind self replicating instructions for future escapes
  • Why LLMs mimic thought but do not think, and why calling it a breach instead of an incident matters legally
  • How the principle of least access applies to AI, and why giving your AI tool delete permissions on a live database is a disaster waiting to happen
  • Why testing AI integrations in a sandbox with read only access first is the only sane approach
  • How pitting two different AI models against each other to check each other's work produces better code than self review