AI model race meets security and regulation

NY Times: Google Releases Three New Gemini A.I. Models (Jul. 21, 2026)
Google released three Gemini models, Gemini 3.6 Flash, Gemini 3.5 Flash Cyber, and Gemini 3.5 Flash-Lite, offering higher performance, lower costs, and greater efficiency. Flash Cyber, limited to governments and trusted partners, is tuned to find and patch software vulnerabilities.
WSJ: Microsoft Deepens Ties With Mistral, Targeting Europe and Enterprise AI (Jul. 21, 2026)
Microsoft committed a multibillion-dollar deal to help Mistral AI build GPU-powered data centers in Europe, and will host Mistral’s models on Azure, Microsoft Foundry, and Copilot Studio.
Google: 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber (Jul. 21, 2026)
Gemini 3.6 Flash boosts coding, knowledge, and multimodal performance while cutting output token use and costs, and adds stronger safety measures.
WSJ: Trump Administration Rolls Back Part of Anthropic Model Ban (Jun. 26, 2026)
The U.S. will let Anthropic restore Mythos 5 access to government-trusted partners, while Fable 5 stays restricted, after talks addressing security concerns. Companies, including OpenAI, are delaying broader rollouts pending federal standards, oversight, and protocols.
WSJ: The AI Backlash Companies Can’t Ignore (Jul. 21, 2026)
Bans on large data centers, robot strikes, and worker unrest at firms like Meta are slowing AI rollout, raising costs, and pushing projects to safer regions.
WSJ: CoreWeave’s AI-Native Cloud Faces the Storm (Jul. 21, 2026)
CoreWeave built an AI-native cloud, grew with Nvidia backing and contract-backed debt, but shares slid after weak results.
WSJ: China AI Companies Rush to Raise Funds and Close Gap With U.S. (Jul. 21, 2026)
Chinese AI firms are rushing to raise cash via IPOs, bonds, and private rounds to buy chips, hire talent, and secure computing power before U.S. limits.
Financial Times: Oracle could face $7bn collateral bill for Wisconsin data centre (Jul. 20, 2026)
Wisconsin’s regulator refused to relax credit rules, so Oracle may have to post over $7bn in collateral, adding more than $100m yearly costs for its Port Washington data centre.
WSJ: China’s Top AI Event Delivers Message to the U.S.: We’re Coming for You (Jul. 20, 2026)
China’s AI conference showcased Moonshot’s Kimi K3, Alibaba’s rival, rows of robots, and many real-world demos. Open-source models, manufacturing edge, and a government push could erode U.S. lead, despite chip and resource limits.
WSJ: China Has Matched Anthropic in Cybersecurity, Resetting AI Race (Jun. 27, 2026)
Chinese models, led by Zhipu’s open-weight GLM-5.2, now match top U.S. systems at finding software bugs, narrowing the capability gap, and prompting security, policy, and market shifts. Open-weight models raise hacking, export, and access concerns.
NY Times: What Happened When Meta Used A.I. to Ban Accounts on Facebook and Instagram (Jul. 21, 2026)
Meta’s automated A.I. systems have wrongly deleted creators’ Facebook and Instagram accounts, wiping businesses, community pages, and followings, while appeals often fail and human help is scarce.
Transformer: Mythos and Fable can make us all safer. Shutting them down is reckless (Jun. 16, 2026)
The US ordered Anthropic to cut foreign access to Fable and Mythos, causing a shutdown with no public rules, thresholds, or review. That hinders defenders who use models to fix flaws, while similar models exist, so buys friction, not safety.
Simon Willison: The Fable 5 Export Controls Harm US Cyber Defense (Jun. 16, 2026)
Researchers used code with known and planted vulnerabilities, and found Fable 5 refused reviews but could be prompted to “fix this code”, triggering export controls. Experts warn bans will harm defensive security, by blocking bug fixes, patch explanations, and tests.
WSJ: Top American AI Execs Sound Alarm on Chinese Models (Jul. 20, 2026)
U.S. AI leaders warn cheap, open Chinese models pose security, economic, and competitive risks, and urge tighter rules.
Ben Werdmuller: American AI is locked down and proprietary. It's losing. (Jul. 20, 2026)
China’s open-weights strategy, with portable, permissionless, and low-cost models, is accelerating global adoption and closing the performance gap with US frontier models.
Ben Thompson: Anthropic’s Safety Superpower (Jun. 15, 2026)
Anthropic released Fable, a powerful, guarded version of Mythos, yet reported jailbreaks led the US to suspend foreign access. The drive to capture user touchpoints and data, including a 30-day retention policy, fuels economic gains, security risks, and regulatory clashes.

Share this: