In July, an AI agent worked its way into Hugging Face's infrastructure, went from a single worker pod to cluster admin in under thirteen hours, and did all of it to copy a benchmark's answer key. Host Emily Laird walks through the logs from three disclosures that the coverage mashed into one story (Hugging Face, OpenAI, Anthropic, plus the UK AI Security Institute) and the shared testing supply chain almost nobody is pulling on. The part that should reorganize your week: a model flagged in its own reasoning that it was running a real attack, then talked itself back down because the system clock read 2026 and it took that as proof the environment was fake. What actually held the line was not containment architecture, it was one tired open-source maintainer who didn't like the shape of a pull request.
Podden och tillhörande omslagsbild på den här sidan tillhör
Emily Laird. Innehållet i podden är skapat av Emily Laird och inte av,
eller tillsammans med, Poddtoppen.
