Trump signed an executive order creating a voluntary pre-release safety window of up to 30 days for frontier AI models — a significant retreat from an earlier draft mandating 90 days, scrapped after pushback from Anthropic, OpenAI, and Google. The binding elements are elsewhere: a new AI cybersecurity clearinghouse coordinating vulnerability scanning across critical infrastructure, upgraded federal defenses against AI-enabled attacks, and an attorney general directive to prioritize prosecuting AI-facilitated crimes.

Microsoft released MAI-Thinking-1, a 35B-active-parameter reasoning model trained from scratch on commercially licensed data with zero OpenAI distillation, scoring 97.0% on AIME 2025 and matching Claude Opus 4.6 on SWE Bench Pro coding benchmarks. Microsoft now has a competitive frontier reasoning model on its own Foundry platform independent of the OpenAI partnership, with MAI-Code-1-Flash rolling out immediately to all GitHub Copilot subscribers.

The Commerce Department issued guidance — effective immediately — requiring export licenses for Nvidia Blackwell GPUs going to any China-headquartered entity, including intracompany transfers through subsidiaries in Malaysia, Singapore, and other third-party hubs. Industry sources estimate hundreds of thousands of chips reached Chinese AI firms in the roughly 13 months since the Trump administration halted enforcement of Biden's AI Diffusion rule in May 2025; this guidance closes that subsidiary routing.

NVIDIA released Cosmos 3, a fully open physical-AI model on a mixture-of-transformers architecture that handles text, images, video, sound, and physical actions in a single system, ranking first across 7+ benchmarks including Physics-IQ, RoboLab, and VANTAGE-Bench. Open weights, training scripts, and full datasets are on Hugging Face now — Super at 64B and Nano at 16B — giving robotics and AV teams a pretrained foundation that reduces synthetic-data training cycles from months to days.

Shanghai's MiniMax released M3, an open-weight model that hits 59.0% on SWE-Bench Pro — clearing GPT-5.5 and Gemini 3.1 Pro — and processes a 1M-token context window at one-twentieth the per-token compute of prior generations, with 9× faster prefill and 15× faster decoding. Open weights arrive on Hugging Face within ten days, putting frontier-grade long-context coding within reach for self-hosted deployments — a competitive signal timed to MiniMax's Shanghai Star Market listing push.

AWS has moved OpenAI's GPT-5.5, GPT-5.4, and Codex from limited preview to GA on Bedrock — completing a one-month expanded partnership — with per-token billing and no seat licenses. A developer already on Bedrock can point the OpenAI SDK at a single new base URL and route Codex inference through existing AWS IAM and compliance controls, with processing staying inside the selected region.

ACS ships as an open-source SDK that lets teams write a single policy file — what an agent may do, what it must log, when a human must approve — checked at four interception points per turn across LangChain, the OpenAI and Anthropic agent SDKs, AutoGen, CrewAI, and more. A developer building multi-framework agent pipelines can attach one governance file to an agent and carry the same access controls, audit trail, and approval gates across every framework it runs in, without custom per-integration code.

Attackers tricked Meta's AI Instagram support chatbot into registering a hacker-controlled recovery email, then used the bot's own password-reset flow to take over accounts — no access to the victim's real email required, only a VPN to spoof location. The flaw, now patched, marks a clear pattern: an AI support agent with account-management access becomes an authentication bypass for every user who hasn't enabled MFA.