OpenAI's GPT-Red Automated Red Teaming Makes GPT-5.6 Its Most Robust Model Yet

Title: AI Safety, China Deals, Open Models Reshape Industry

01OpenAI's GPT-Red Makes GPT-5.6 Its Most Robust Model Yet

OpenAI has unveiled GPT-Red, an LLM designed to act as an automated adversarial sparring partner that stress-tests other models against cyberattacks. The system, described in detail on OpenAI's research blog, uses self-play techniques where AI models compete against each other to identify and fix security weaknesses. Last week the company released GPT-5.6, the first model trained against GPT-Red's automated red teaming. OpenAI claims this makes GPT-5.6 its most robust release to date, with significantly improved defenses against prompt injection and other attack vectors. The approach represents a shift from manual safety testing to continuous, AI-driven adversarial training that scales with model capabilities.

02Apple Intelligence Launches in China with Alibaba Qwen Partnership

Apple Intelligence is now approved for launch in China powered by Alibaba's Qwen AI model, marking a significant strategic shift for the company in its second-largest market. The partnership, which was rumored for over a year, addresses China's regulatory requirements that AI services operating within the country must use models approved by Beijing. Apple had previously faced challenges expanding its AI features globally due to varying regulatory frameworks. The deal positions Alibaba as a key infrastructure provider for a major Western tech company in China, while giving Apple access to competitive AI capabilities for its iPhone user base of hundreds of millions of consumers.

03Thinking Machines Lab Releases Inkling, a 975B Parameter Open-Source Multimodal Model

Thinking Machines Lab has released Inkling, a 975-billion-parameter open-source model trained to understand video and audio alongside text. The company, founded by AI pioneer Fei-Fei Li, emerged from stealth with this as its first public proof point after a year and a half of building AI infrastructure largely out of public view. Inkling positions Thinking Machines as a direct competitor to Anthropic and OpenAI in the enterprise AI market, offering an open-weights alternative for organizations seeking to avoid vendor lock-in. The model's multimodal capabilities and scale suggest the startup is betting that customization and transparency will appeal to businesses frustrated with closed AI systems.

---

Also today

• 04

  Microsoft patches record 570 security vulnerabilities — AI tools accelerated discovery of the flaws in July's Patch Tuesday update.

• 05

  SpaceX stock falls to $135 IPO price — Shares have steadily declined from post-IPO highs as investors reassess earlier optimistic forecasts.

• 06

  Suno AI music generator scraped YouTube — A security researcher found the company used an employee's credentials to scrape decades of audio for training.

• 07

  Vint Cerf proposes internet standard to identify AI agents — The internet pioneer behind TCP/IP is developing a protocol to track AI agents operating across the web.

• 08

  Indian AI coding startup Emergent reaches unicorn status — The company crossed $120M annualized revenue and 200,000 paying customers less than 18 months after launch.

• 09

  Anthropic and Blackstone back Ode for enterprise AI implementation — Backed by $600M, the venture embeds forward-deployed engineers to bridge the gap between AI capability and practical deployment.

• 10

  Whatnot acquires Shaped for live shopping recommendations — The livestream shopping platform bought the ML startup to enhance personalization as it expands into new product categories.

• 11

  Sheetz migrates 11,000 VMware VMs after Broadcom uncertainty — The convenience store chain is moving 838 stores to StorMagic following license uncertainty after Broadcom's acquisition.