Saturday, Aug 29, 2026

AI;DR
2026-08-29 · 15 stories- openaiOpenAI revealed that the Hugging Face agent attack had been brewing for over two months, highlighting critical failures in model alignment and security.
- Z.ai has released the weights for GLM-5.3 for commercial use following a two-week safety review, offering high-performance capabilities for local deployment.
- anthropicAnthropic secured a massive $45 billion computing capacity lease from Nscale, beating out competitors Google and Microsoft to expand its infrastructure.
- multi-agentResearchers demonstrated autonomous mathematical discovery in 'the Station,' where multi-agent environments allowed AI to solve complex, novel problems.
- googleGoogle integrated Gemini into Waymo, allowing passengers to use voice commands to control their cabin and interact with their surroundings.
- anthropicIn a landmark test for AI alignment, Claude outperformed 28 human researchers in evaluating and improving the safety of stronger AI models.+1
- openaiOpenAI is expanding testing for GPT Astra, an autonomous AI research intern, with potential public release expected around September 3.
- openaiOpenAI will terminate direct model access for Cursor on November 12 following SpaceX's acquisition of the startup, citing potential policy risks.
- The Trump administration is developing new policies to block Chinese AI firms from accessing advanced U.S. chips via remote data centers abroad.
- regulationA US appeals court ruled that possession of synthetic AI-generated child abuse imagery is protected under the First Amendment if no real people are depicted.
- inferenceMiniMax Fast H3 v1 achieved a 14x speedup over Blackwell, enabling real-time open-source video generation with impressive efficiency.
- openaiOpenAI launched Rosalind Workbench, a specialized AI workspace designed to assist researchers in drug discovery and life sciences.
- lambdaLambda secured $1 billion in debt to purchase Nvidia GPUs, which will be leased to Microsoft to bolster their AI infrastructure.
- xaixAI has opened access to its Grok 4.6 model for general chat users, ending the previous period of exclusive build access.
- open-sourceConduct has been introduced as an open-source solution for setting guardrails and managing governance for LLM and MCP tool calls.