Episode Details

Security and accessibility take center stage as David Kuszmar and Carnegie Mellon’s SEI CERT disclose universal jailbreak vulnerabilities affecting models from OpenAI, Google, Anthropic, and Meta. To combat these risks, OpenAI has unveiled GPT-Red, a specialized LLM used for automated red-teaming that outperformed human testers in identifying prompt injection flaws. The episode also explores the darkening social atmosphere for AI leaders, with reports from the Wall Street Journal detailing a surge in threats and physical security incidents at major labs. In model news, Mira Murati’s Thinking Machines has launched Inkling, a 975-billion-parameter multimodal system under an Apache 2.0 license, while South Korea moves toward national AI sovereignty by providing free access to its entire population.

Topics Covered

• 🔬 Systemic LLM exploits: Time Bandit and Inception vulnerabilities.
• 🛡️ GPT-Red: OpenAI’s new automated super-hacker for safety.
• ⚠️ Security backlash: The rising physical threats to AI executives.
• 🤖 Inkling release: Thinking Machines’ multimodal open-source entry.
• 💻 GLM-5.2 and coding: A new leader in open-weight performance.
• 🌐 National AI: South Korea's free access program for all citizens.

Neural Newscast is AI-assisted, human reviewed. View our AI Transparency Policy at NeuralNewscast.com.

• (00:12) - Introduction
• (00:12) - Systemic Vulnerabilities and GPT-Red
• (00:12) - Physical Security and The Executive Backlash
• (00:44) - New Model Releases: Inkling and GLM-5.2
• (00:44) - Global Infrastructure: South Korea and Bengaluru