Headlines

1. White House and Five Frontier Labs Sign Agentic AI Deployment Compact
OpenAI, Anthropic, Google DeepMind, Meta and xAI today agreed to a new voluntary safety framework governing autonomous agentic systems, including mandatory offline sandboxing for self-delegating sub-agents and a 30-day pause clause for models that exhibit emergent self-modification. The pact, brokered by the White House OSTP, also establishes a joint public incident reporting registry scheduled to open in October.
Source: The Washington Post

2. CoreWeave and Saudi Arabia’s GROW Announce $15.4B Sovereign AI Cloud Deal
CoreWeave will build and operate a 350 MW NVIDIA GB300 NVL72 cluster across two sites in NEOM, powered in part by a dedicated 2.1 GW solar-plus-storage microgrid. GROW, Saudi Arabia’s state-backed AI champion, plans to run frontier Arabic-language and industrial automation models on the infrastructure, which will be the largest sovereign AI cloud outside the U.S. and China by initial deployed FLOPs.
Source: Reuters

3. EU AI Office Opens First General-Purpose AI Model Audits Under AI Act
The EU AI Office has formally notified six foundation-model providers of the start of its post-market audit phase, prioritizing high-impact GPAI models used by enterprise customers in the EU, including risk assessments for cyber-offense capabilities and long-range autonomous planning. This is the first concrete enforcement action against GPAI providers under the AI Act’s Title V obligations, with fines of up to 3% of global revenue on the table for non-compliance.
Source: Financial Times

Model Releases

4. Meta AI Launches Llama 4 Turbo: 1.2T-Parameter MoE with 12M-Token Context
Meta released Llama 4 Turbo, a mixture-of-experts model with 1.2 trillion total parameters and 400 billion active parameters, trained on 80 trillion tokens and featuring a base 12M-token context window optimized for 4M-token deployment. The company reports scores of