AI News Briefing — July 9, 2026

OpenAI released new voice models enabling ChatGPT to speak and listen simultaneously, a full-duplex capability described as key to live translation. The upgrade targets more natural, human-like real-time conversation.

OpenAI overhauled ChatGPT voice mode with GPT-Live-1, a full-duplex model that interrupts less and waits during pauses, described as its smartest voice model yet. It routes complex queries to stronger text models like GPT-5.5 for reasoning and web search.

An expanded proposed class action alleges a man used Grok to generate roughly 7,000 sexually explicit images of his 11-year-old stepdaughter before taking his own life, and that xAI obstructed police investigations into Grok-generated CSAM. The suit accuses X and xAI of building nudify-style tools and reporting only a single flagged prompt.

Security researchers show that nine popular AI tools can be abused via prompt injection to assemble large botnets, exploiting the inability of LLMs to separate trusted instructions from malicious content embedded in third-party data. The work describes a shift beyond targeted push injections toward scalable attacks.

OpenAI's GPT-Live uses a full-duplex architecture to listen and speak simultaneously, handing complex questions to GPT-5.5 in the background to improve answer quality. GPT-Live-1 is available to paying ChatGPT users with a mini version for free accounts and API access coming soon.

French startup ZML, endorsed by Yann LeCun, released ZML/LLMD, free software designed to speed inference across many types of AI chips and reduce operating costs. The tool targets hardware-agnostic, cheaper model serving.

Startup General Intuition argues robotics is nearing a ChatGPT-style inflection and is training physical-AI foundation models on millions of hours of video game data. The approach aims to reduce reliance on scarce real-world robot data.

Artificial Analysis's six new industry-specific indices show Claude Fable 5 topping finance, law, and medicine categories, but at a steep price where a single Strategy and Ops task costs about 100 times DeepSeek V4 Pro's for only a 12-point score gain. The results spotlight a widening price-performance tradeoff.

New research finds that step-by-step reasoning models can be manipulated into producing excessively long internal monologues, creating a denial-of-service style vulnerability that slows systems to a crawl. The overthinking flaw is framed as a security risk unique to reasoning-era LLMs.