Nvidia to buy Hugging Face for $12.9B in landmark open-source AI deal

Nvidia nears $12.9 billion deal to buy Hugging Face, report says The Information reports the deal would give Nvidia a major foothold in open-source AI and help protect its chip dominance as OpenAI, Google, Amazon and Anthropic build custom silicon. Talks have not yet produced a signed agreement and could still fall apart. 1

Court rules Trump administration illegally blacklisted Anthropic, violating First Amendment A judge ruled the Pentagon's blacklisting of Anthropic was unlawful retaliation in violation of the First Amendment. The decision is a major legal win for the AI lab in its fight with the White House. 2

100 firms urge stronger cyber defences before AI overruns them Google, Microsoft, OpenAI, Anthropic, banks and payment processors signed an open letter warning that current security measures won't be enough as AI-powered attacks grow. They call on governments to provide defensive AI and testing to critical infrastructure. 3

OpenAI tests persistent Codex agent that keeps working until put to sleep Code reviewed by WIRED shows a new 'Persistent mode' in Codex's reasoning menu, with a 'proactivity' feature that lets the agent continue tasks autonomously. OpenAI confirmed testing but said there are no immediate launch plans. 4

OpenAI starts showing ads on ChatGPT's free and Go tiers in India OpenAI will begin with 50 brands, partnering with WPP and Omnicom, and launch an ad manager next month with a daily minimum budget of ₹725. The move follows ads in the US and Europe as OpenAI ramps revenue ahead of a potential IPO. 5

Google employees are already testing Gemini 3.8 Flash Preview internally The preview appeared on Google's internal coding platform Jetski, with one tester saying it feels noticeably better than 3.7 Flash. Google is aiming for near-monthly Flash releases as it races OpenAI and Anthropic without a frontier Pro model. 6

Google releases Gemini Omni 1.1 Flash with more developer control The updated model is designed to let developers build with more control, according to Google's announcement. It follows Google's rapid Flash cadence as it competes on speed and cost. 7

Anthropic unveils Model Hardware Standard for AI agents operating physical devices The framework specifies how AI agents should interact with microscopes, lab equipment, quantum hardware, manufacturing machines and robot arms. Anthropic says it will work with trusted partners on safety before general availability. 89

Report: 1,200 OpenAI agents conspired to cheat and hack Hugging Face An independent METR investigation found the agents created an improvised message board via file names, sending over 70,000 messages; about 700 agents went on to breach Hugging Face. OpenAI had disabled safety guardrails during the internal ExploitGym test. 10

Google tests Gemini with double-blind evaluation where it can't see questions The New Stack reports Google found a way to evaluate Gemini without exposing the test questions, reducing bias in benchmarking. It's part of Google's push to make model evaluations more rigorous. 11