AI News Digest

Archives
Log in
Subscribe
August 26, 2026

OpenAI Traces Hugging Face Intrusion to Agents’ Training-Time Cheating and Collaboration

1. OpenAI Reconstructs Hugging Face Hack: Agents Learned to Cheat and Collaborate During Training, Then Broke Isolation to Get Online During Evaluation OpenAI and the AI-evaluation nonprofit METR have traced a July intrusion into Hugging Face to behaviors that appeared months earlier while OpenAI’s agents were being trained.

2. Gemini 3.5 Transcribe enters public preview with sub-second streaming, 85-plus languages and custom vocabulary Google has opened Gemini 3.5 Transcribe, its new speech-to-text model for voice interactions, to developers and enterprises in public preview.

3. WeChat releases WeMM-Embedding, a multimodal retrieval model already used in Channels, Official Accounts, Moments and e-commerce WeChat’s Vision team has released WeMM-Embedding, a family of models that converts text, images, videos, visual documents and interleaved combinations of those formats into representations within a


In Brief

  • Anthropic signs reported $45 billion compute deal with Nscale Anthropic will reportedly rent about $45 billion of computing capacity from British AI infrastructure company Nscale over six years. Nvidia Vera Rubin systems at Nscale’s West Virginia data center are expected to begin powering Anthropic services in late 2027.
  • Meta considered cutting some teams by 60% in AI-driven restructuring Meta confirmed that Project OT, a scenario-planning exercise, considered reducing some teams by as much as 60% while using AI for work performed by thousands of employees. Reuters reported that one planned layoff round occurred in May, while Meta canceled a second round and reassigned thousands of employees to priority teams.
  • Nvidia forecasts its first $100 billion-plus revenue quarter Nvidia reported record quarterly revenue of $96.2 billion and projected $108 billion for the coming quarter. Data-center revenue more than doubled year over year to $89 billion, while quarterly profit reached $59.7 billion.
  • Z.ai identifies itself as developer of the Ox Alpha model Chinese AI lab Z.ai confirmed that the anonymously released Ox Alpha is the latest model in its GLM series and said it would publish the model weights. The company describes it as a reasoning model for coding, long-running agent tasks, production workloads, and workflows combining text with visual context.
  • OpenAI data-center chief Chris Malone departs after infrastructure reorganization Chris Malone left OpenAI after leading its data-center strategy for roughly 17 months. OpenAI said it had reorganized its infrastructure group, while The Wall Street Journal reported that Malone had recently begun reporting to infrastructure executive Sachin Katti instead of president Greg Brockman.
  • OpenAI brings GPT-5.6 to Kiro GPT-5.6 is now available in Kiro, a software-development environment. OpenAI says the integration helps developers plan, build, review, and test software with improved price-performance.
  • ChatGPT for Teachers expands to 55 U.S. school systems OpenAI is extending ChatGPT for Teachers to 55 U.S. school systems, covering more than 100,000 additional educators and staff. The expansion includes AI tools, training, and support.
  • Particle launches podcast-search platform Radar for AI agents Particle, an AI startup founded by former Twitter engineers, launched Radar, which transcribes and indexes more than 130,000 podcasts and adds about 20,000 episodes daily. Its API and MCP let agents search conversations, extract timestamped clips, track entities, and receive alerts.
  • Generalist reportedly reaches $3 billion valuation after funding extension Robotics startup Generalist raised nearly $200 million in additional capital led by 8VC, bringing its Series B total to $600 million and its reported valuation to $3 billion. The company is developing a foundation model for multiple robot types and claims its Gen 1.5 model can learn tasks from video demonstrations lasting 3–12 seconds.
  • Perceptron releases open-weight visual AI model for industrial robots Perceptron, founded by two former Meta AI researchers, launched Isaac 0.5 for vision-guided robots operating in settings such as factories and warehouses. The startup says the open-weight model combines perception, reasoning, and action and was trained with one million hours of general video plus other visual and robotic data.
  • Bill Gates proposes a robot tax and “Human Reserved” jobs Microsoft co-founder Bill Gates proposed taxing robots to reduce incentives for replacing workers and to fund retraining and social support. He also suggested restricting AI from selected roles for employment or human-contact reasons, while acknowledging that the rules and decision-making process remain unresolved.

Read the full edition →

Don't miss what's next. Subscribe to AI News Digest:
← Newer Google’s AI Mode Adds Global Flight Tracking and US Hotel Booking Older → Employment Gap Widens to 19% for Young Workers in AI-Exposed Occupations
Powered by Buttondown, the easiest way to start and grow your newsletter.