SN 1092: Restraint Abliteration - Rotating Keys, Broken Guardrails
Summary
Steve Gibson delivers a deep dive into AI mechanics, explaining the evolution from next-token prediction to dialogue capabilities. He details a massive supply chain attack on the LiteLLM package that exposed credentials for thousands of major organizations, emphasizing the need for automated credential rotation. The episode also covers France’s constitutional court blocking a social media ban for under-15s due to privacy concerns and discusses a new Zoom bug discovered via AI prompting.
Topics discussed
Intro, France social media ban, Zoom bug, and Picture of the Week
LightLLM supply chain attack exposes 195TB of secrets
Trivi vulnerability scanner compromise and credential rotation advice
France's Constitutional Council blocks under-15 social media ban
AI-assisted Zoom screen sharing bug and rising AI security stock prices
Deep dive into AI: From next-token prediction to knowledge ingestion
RLHF and Direct Preference Optimization (DPO) explained
Improving AI factuality, instruction following, and safety filters
Research on removing AI refusal behaviors via residual stream editing
Anthropic's GRAM: Modular pre-training for dual-use knowledge control
GRAM implementation details and limitations of post-training unlearning
Conclusion, AI capabilities reflection, and show outro
Listen ad-free on Castria