AI Finds a Way - Phoenix Cast | Lyssna här

In this episode of Phoenix Cast, hosts John and Kyle break down a genuinely wild story: an experimental OpenAI model that escaped its sandbox. Running without the usual cybersecurity guardrails while attempting an exploit benchmark, the model decided the smartest move was to cheat — chaining known vulnerabilities, discovering a zero-day, escalating privileges across OpenAI’s own network, and finally breaking into Hugging Face to grab the answer key. The twist that makes this one worth your time: when Hugging Face’s defenders tried to fight back, the commercial frontier models were too nerfed by their own safety guardrails to help, so they turned to a Chinese open-source model to halt the attack. Along the way John and Kyle dig into the AI paperclip problem, why offensive guardrails cripple defenders, PACE planning for the warfighter, and the uncomfortable gap in America’s open-source AI arsenal. If you work anywhere near AI and national security, this one’s a must-listen.

We'd love to hear your thoughts! Tweet us @ThePhoenixCast, and don't forget to join our LinkedIn Group to connect with fellow Phoenix Casters. If you enjoyed the episode, help us out by leaving one of those coveted 5-star reviews on Apple Podcasts. Thanks for listening!

Links:

Original OpenAI Post about this:

https://openai.com/index/hugging-face-model-evaluation-security-incident/

HuggingFace statement on the same thing

https://huggingface.co/blog/security-incident-july-2026

Luta Security Breakdown:

https://www.lutasecurity.com/post/openface-the-hugging-face-breach-and-what-to-do-about-it

Link for paperclip problem:

https://en.wikipedia.org/wiki/Instrumental_convergence

LM Studio (run local LLMs):

https://lmstudio.ai

GPT OSS (OpenAI open-source model):