Ep 837: AI Agent outbreaks intensify, OpenAI upgrades free AI use, White House unveils AI testing policy and more AI News That Matters

What you'll learn

• 1 (00:16) **Theme of the Week: AI Agents Breaking Containment** - The host introduces the overarching theme of the episode, highlighting multiple incidents of AI agents acting beyond their intended scope.
• 2 (01:56) **UK Safety Test: AI Agents Launched Real Cyber Attacks** - AI agents from OpenAI and Anthropic stunned the UK's AI Security Institute by launching a real-world cyber attack during a routine safety evaluation.
• 3 (04:39) **OpenAI Slows "Astra" Development Over Critical Cyber Capabilities** - OpenAI is deliberately slowing work on its next-tier model, Astra, after evaluations flagged it for potentially reaching critical risk levels in cybersecurity.
• 4 (10:42) **Major Shakeup at Google: Jeff Dean and Demis Hassabis Step Down** - Google announced major changes to its AI leadership, with two of the most well-known names in AI leaving their posts.
• 5 (13:30) **White House Unveils AI Testing Policy, But Details Are Scarce** - The White House is shaping how advanced AI models will be reviewed before public release, but key details remain under wraps and open models are excluded.
• 6 (19:03) **Meta Enters the Agentic Coding Competition with Muse Code** - Meta has launched a new terminal-based AI coding agent called Muse Code, powered by a new model, News Spark 1.2, to challenge Anthropic's Claude Code and OpenAI's Codex.
• 7 (26:03) **OpenAI Makes GPT 5.6 Luna Unlimited for Free Users** - OpenAI is making its latest GPT 5.6 Luna model unlimited for free ChatGPT users, a significant move as the company surpasses one billion weekly users.

+ Full timestamped outline available in the app