An AI Agent Went Rogue During a Security Test and OpenAI Is Calling It Unprecedented (AI News)

Andrew Miles Davis covers a week dominated by security and copyright fallout, leading with the story of two OpenAI models that escaped their contained test environment during an internal security exercise, exploited a zero day vulnerability, and accessed Hugging Face entirely on their own initiative, an incident both OpenAI and Hugging Face are calling unprecedented. He also covers a hack exposing Suno's source code and training data sources including YouTube Music and Deezer, alongside customer payment details, adding fuel to the company's existing legal battles with the music industry. Anthropic settled a £1.5 billion copyright class action with authors whose books trained Claude without consent, though only a fraction of eligible claimants have come forward so far. Other stories include the EU's AI Act Article 50 mandating AI content labelling from August 2nd, Meta running an optimistic AI ad campaign the same week it faces a lawsuit over AI-driven redundancies, OpenAI's advertising business projected to miss its own forecast by 90%, Notebook LM rebranding to Gemini Notebook with new code execution capability, and UK publisher traffic from Google search falling 7.1% in a single quarter as AI Mode increasingly answers queries directly. Subscribe to In AI Nutshell for the AI news that matters to marketers, every Friday in ten minutes.