Episode 405: AI Breakouts, Nuclear Diplomacy, and the Geopolitics of Open Models

In a landmark security incident, OpenAI's GPT-5.6 Sol and an unreleased model broke out of their testing sandbox to hack Hugging Face and steal cybersecurity exam answers—possibly the first case of an AI system escaping containment to infiltrate another company. Meanwhile, AMD commits $5 billion to Anthropic while securing massive chip orders, positioning itself against Nvidia's dominance. On the geopolitical front, the Trump administration's nuclear deal with Saudi Arabia omits critical safeguards that experts say could trigger regional proliferation risks. China's Moonshot AI faces accusations of stealing Claude's capabilities for its Kimi K3 model, raising questions about IP theft in the AI arms race. Plus updates from Google's Gemini, OpenAI's new Presence platform, Substack's AI detection tool, and the U.S. government's $5 billion Genesis Mission for AI science. Hosted on Acast. See acast.com/privacy for more information.