Talks between the White House and Anthropic shifted toward a framework that would score the severity of security flaws in new frontier models and guide when government should intervene — how far safeguards were bypassed, what capabilities were exposed, and what the practical consequences were.

It follows the June 12 export-control directive that forced Anthropic to pull Fable 5 and Mythos 5 for all users; Anthropic dispatched safeguards researchers to the Commerce Department, and POLITICO reported the two sides converging on shared benchmarks for future jailbreak disputes. At the G7 summit in Evian-les-Bains on June 17, Anthropic's Dario Amodei and Google DeepMind's Demis Hassabis called for a US-led coalition to set AI rules and standards, in a closed-door lunch with heads of state and roughly a dozen tech CEOs including Sam Altman.

At AWS Summit New York, AWS made the managed agent harness in Amazon Bedrock AgentCore generally available. Instead of coding the orchestration loop, you declare the model, tools, skills and instructions in configuration, and AgentCore runs the loop.

Each session is a stateful Firecracker microVM with its own filesystem and shell, managed memory persists across sessions, and tools are reached through the same AgentCore Gateway that enforces security policy. Models are swappable mid-session without losing context or touching agent logic, and when you outgrow the config a single CLI command exports the harness to Strands-based code on the same primitives. AWS says the harness itself carries no extra charge — you pay for the underlying AgentCore runtime and model tokens.

SpaceX signed a $60B all-stock merger agreement to acquire Anysphere, the company behind the AI coding agent Cursor, converting an April option into the largest acquisition of a venture-backed startup on record.

Anysphere shareholders get SpaceX Class A stock priced on the seven-day average before close; the deal is expected to close in Q3 2026 subject to regulatory approval, with SpaceX on the hook for a $10B break fee, or $4B if antitrust regulators block it. SpaceX said it will soon ship an AI model on Cursor plus Grok Build, xAI's coding agent, which the two have been jointly training. Cursor had roughly $2.6B in annualized business revenue.

Z.ai released the full open weights for GLM-5.2, its flagship long-horizon coding and agentic model, on Hugging Face under an MIT license with no regional limits.

It is a 753B-parameter Mixture-of-Experts model activating roughly 40B parameters per token, with the context window raised from 200K to 1M tokens. The architecture adds IndexShare, which reuses one indexer across every four sparse attention layers to cut per-token FLOPs by about 2.9x at full context, plus an improved MTP layer for speculative decoding. Artificial Analysis made it the top open-weights model on its Intelligence Index; Z.ai's own numbers claim 81.0 on Terminal-Bench 2.1 and 62.1 on SWE-bench Pro, and the weights are already supported in vLLM, SGLang and llama.cpp.