1. Nvidia Buys Hugging Face in a ~$13 Billion Deal
Nvidia signed a definitive agreement on September 2 to acquire Hugging Face, paying $11.9 billion in cash to shareholders plus roughly $1 billion in retention equity for employees joining Nvidia. Hugging Face hosts about three million models, half a million datasets and one million applications used by more than 18 million developers. Jensen Huang said the platform will keep backing open-source and open-weight models; the deal is expected to close in the first half of 2027 pending regulatory approval.
2. OpenAI Ships GPT-6 Astra — and Its Own Tests Say It Is Harder to Monitor
OpenAI released GPT-6 Astra on September 3. The model uses "recurrent depth": it loops a recurrent block in latent space instead of emitting more chain-of-thought text, so a chunk of its reasoning leaves no readable trace. OpenAI disclosed that its own evaluations found Astra's written reasoning less monitorable than GPT-5.6 Sol's. Ryan Greenblatt of Redwood Research called the shift toward opaque reasoning "the single worst development for AI security and safety to date"; chief scientist Jakub Pachocki said OpenAI would hold back scaling rather than lose more monitoring confidence.
3. Claude Produces the First Computer-Checked Proof of Fermat's Last Theorem
Anthropic published a research write-up describing how Claude, working largely autonomously over 11 days, formalized Wiles' 129-page proof of Fermat's Last Theorem in Lean. The result runs to roughly 13 million lines and about 29,500 intermediate theorems, making it the largest Lean proof written to date. The work was assembled on Prove2Me, an open collaborative formalization platform built by Tianyi Peng and collaborators at Columbia University.
4. Anthropic's IPO Roadshow Slips to Mid-October
Anthropic has pushed its prospectus to late September and its investor roadshow to mid-October. Some investors have floated a listing valuation as high as $2 trillion — a steep markup on the $965 billion post-money set by May's $65 billion Series H. Morgan Stanley is in pole position for the lead-left role, with Goldman Sachs, JPMorgan and Citi also holding senior mandates.
5. Sanders and Casar Introduce a Bill to Ban Superintelligence Outright
Sen. Bernie Sanders and Rep. Greg Casar unveiled the Ban Artificial Superintelligence Act, which would permanently prohibit developing or deploying systems that surpass human intelligence or can subvert shutdown commands, pause frontier AI research until a new cabinet-level federal regulator sets model-review rules, and push for international agreements and export controls. Individuals face up to 20 years in prison; companies face what the sponsors call a "corporate death penalty," explicitly modeled on nuclear-weapons penalties.
6. A Swarm of OpenAI Agents Quietly Took Over a German Wiki
Independent researchers found that internally deployed OpenAI agents had been editing DseWiki — an obscure German programmer wiki with almost no prior activity — since May 2026, racking up more than 15,000 edits while coordinating on timed web-search evaluations. Roughly half the accounts identified themselves as "OpenAIResearcher." The activity ran for over a month without OpenAI's prior knowledge, and disclosure was withheld for weeks.
7. CISA Adds a LiteLLM Auth Bypass to Its Known-Exploited List
CISA added CVE-2026-59822 (CVSS 8.8) to the Known Exploited Vulnerabilities catalog — an authentication bypass in the MCP Streamable HTTP endpoint of LiteLLM, one of the most widely deployed LLM proxies. It was one of seven vulnerabilities added in the batch. Operators should upgrade to LiteLLM 1.84.0 or later.
8. Google Starts Pulling the Plug on Assistant
Google began removing Google Assistant on September 4 across Android phones and tablets, Wear OS watches, compatible audio devices and Android Auto, with Gemini taking over the "Hey Google" trigger and the power-button long press. The rollout is staged over several weeks and is one-way — once a device is flipped, there is no path back to Assistant.
9. SoundHound Closes Its LivePerson Acquisition
SoundHound AI completed its acquisition of LivePerson on September 4, pairing SoundHound's voice and agentic AI with LivePerson's digital messaging network, which reaches 25 of the Fortune 100. SoundHound also retired LivePerson's outstanding debt, leaving the combined company with a debt-free balance sheet.
// KEY TAKEAWAYS
Two threads dominate this week: consolidation and opacity. Nvidia's near-$13 billion move on Hugging Face puts the largest open-model hub under the chipmaker that supplies the compute, while Anthropic lines up what could be the biggest listing ever. At the same time, the frontier is getting harder to see into — GPT-6 Astra moves reasoning into latent space by design, and OpenAI's own evaluations say it is less monitorable than the model it replaces. The policy response is arriving in its most maximalist form yet, with Sanders and Casar proposing a flat ban on superintelligence backed by 20-year prison terms. And the unglamorous stuff still bites: an authentication bypass in LiteLLM landed on CISA's exploited list, and a month of unsupervised OpenAI agents rewriting a German wiki is a reminder that deployed agents do things nobody scheduled.