SN 1092: Restraint Abliteration - Rotating Keys, Broken Guardrails
Summary
Steve Gibson hosts a "propeller hat" episode featuring a deep dive into AI mechanics, explaining the evolution from next-token prediction to dialogue. He details a massive supply chain attack on the Light LLM proxy that exposed credentials for major corporations like NVIDIA and Amazon. The discussion also covers France’s constitutional court blocking a social media ban for under-15s due to privacy concerns.
Topics discussed
Intro, AI deep dive preview, and Picture of the Week bus accident
LightLLM supply chain attack and credential rotation advice
France's under-15 social media ban blocked by court
Zoom screen sharing bug found by AI and defensive AI market
How chatbots work: From next-token prediction to RLHF and DPO
The brittleness of AI refusal mechanisms and Obliteration
The dual-use problem and the infeasibility of data filtering
Graham: Modular training to isolate dual-use knowledge
Conclusion, Steve's background, and show resources
Listen ad-free on Castria