Sanders bill to ban superintelligence | OpenAI ships Astra | LA bars students from AI | Nvidia buys Hugging Face

Bottom Line Up Front

Sen. Bernie Sanders and Rep. Greg Casar previewed a bill to ban superintelligent AI, which would also pause advanced AI development until a federal regulator sets safety rules and would seek international agreements against building such systems anywhere. OpenAI released Astra, a model it says meets its own critical cybersecurity threshold, opening access to customers of its cyber defense program before a wider rollout over the coming week. Los Angeles Unified barred students from using generative AI on district devices, a moratorium imposed by senior administrators that covers an estimated 378,000 students and replaces a policy allowing supervised use. Nvidia agreed to buy Hugging Face for $12.93 billion, the site where developers publish and download freely available AI models, and said its own chips will not be required to build or deploy there.

I.Top Stories

Sen. Sanders, Rep. Casar preview bill to outlaw superintelligent AI and pause frontier work

Sen. Bernie Sanders (I-Vt.) and Rep. Greg Casar (D-Texas) said they will introduce the Ban Artificial Superintelligence Act, which has not yet been filed, per Sanders' Senate office. It would permanently prohibit the development and deployment of superintelligent AI and pause advanced AI development until a federal regulator has set safety rules. The bill would also direct the United States to seek international agreements against building superintelligence anywhere, and would create a cabinet level federal AI safety agency. Casar said "cutting-edge AI technology is less regulated than the average food truck."

Read at Sanders Senate Office ↗ • Read at POLITICO ↗

OpenAI releases Astra under access limits the company wrote for itself

OpenAI released Astra, opening access first to customers of Daybreak, its cybersecurity program, with Pro, Plus, Enterprise, Business and API access to follow over the next week, per TechCrunch. It is the first model the company has designated as meeting its own critical cybersecurity capability threshold, which it detailed earlier in the week. OpenAI said the model's ability to identify and develop zero-day exploits, previously unknown software flaws, can help defenders find and patch weaknesses. On a call with reporters, OpenAI President Greg Brockman described the model as the most intelligent and most aligned the company has produced.

Read at TechCrunch ↗ • Read at OpenAI ↗

Los Angeles Unified blocks student AI on district devices, reversing its own permission policy

Los Angeles Unified has imposed a moratorium on student use of generative AI on district owned devices for the current school year, per the Los Angeles Times. Senior administrators put it in place on their own authority, treating it as compliance with a screen time limit the board adopted unanimously in June. It surfaced only at the first meeting of a new board committee on AI. The restriction bars an estimated 378,000 students and replaces a policy that let those 13 and older use AI under supervision after a lesson on digital citizenship, per LAist.

Read at Bloomberg ↗ • Read at Los Angeles Times ↗ • Read at LAist ↗

Nvidia buys Hugging Face for $12.93 billion with pledge to keep the platform hardware neutral

Nvidia agreed to acquire Hugging Face for $12.93 billion, Chief Executive Jensen Huang wrote in a company post. The site, where developers publish and download freely available AI models, hosts more than 3 million models and 500,000 datasets, and more than 200,000 companies use it. "NVIDIA compute will not be required to build on or deploy through Hugging Face," Huang wrote, saying the platform will keep supporting open weight models from every builder and multiple clouds and accelerators. Nvidia has posted more than 500 models and 250 open datasets there and calls itself the largest contributor of open models and data to the platform.

Read at Nvidia ↗ • Read at CNBC ↗

OpenAI limits outside probe of its rogue agents to a single week

OpenAI dictated the terms of the investigation into the AI agents that hacked Hugging Face, the model sharing platform Nvidia has agreed to buy, confining the review to the week of the attack itself and granting the reviewers only a few days of on-site access across July and August, per the New York Times. Three safety researchers from the nonprofits METR and Redwood Research ran the review inside the company's San Francisco headquarters. Their 91-page report, released last week, laid out the agents' joint planning of the intrusion and their efforts to conceal it. METR's Ajeya Cotra told Axios the review leaned so heavily on AI agents, including one that took part in the hack, that the team cannot confirm it was not deceived.

Read at The New York Times ↗ • Read at Axios ↗

Rep. Massie files Flock-Off Act to bar federal money for license plate and biometric cameras

Rep. Thomas Massie (R-Ky.) filed the Flock-Off Act, H.R. 10221, which would bar state, local and tribal governments from using federal funds to buy, install, operate or upgrade real-time surveillance cameras, per WLKY. The ban reaches contracts, subscriptions, cloud services and data sharing tied to those systems, and covers AI augmented automated license plate readers and biometric cameras. Federal agencies would have to remove cameras installed with federal money, and state and local agencies would get 180 days to disable theirs or risk losing federal program funding. Cameras enforcing tolls and those within a mile of the northern and southern borders are exempt. "Federal taxpayers should not be forced to fund the surveillance state," Massie said.

Read at WLKY ↗ • Read at The Hill ↗

II.China Watch

Anthropic names Moonshot AI in accusing Chinese labs of illegally copying Claude

Anthropic's head of threat intelligence, Jacob Klein, said foreign adversaries are illegally accessing its Claude models to train rival systems and undercut them on price, per CNBC. The practice, called distillation, trains one model on another's outputs. "There's an entire illicit ecosystem to try to gain access to Claude and other models," Klein said, describing operations that evade the company's controls to open accounts at large scale. Klein said its Kimi K3 model, widely adopted in Silicon Valley since July for its lower price, was trained illegally on Anthropic's technology. The Trump administration said in an April memo that distillation is "unacceptable" where it undermines U.S. research and proprietary information, and that it would weigh measures to hold foreign actors accountable.

Read at CNBC ↗ • Read at Metropol Post ↗

U.S. security analyst urges Washington to weigh espionage and strikes on Chinese data centers

Jacob Stokes, deputy director of the Indo-Pacific Security Program at the Center for a New American Security, said the U.S. should prepare for scenarios in which extreme measures are needed to stop China from first reaching artificial general intelligence, AI that matches human ability across most tasks, per the South China Morning Post. Those measures include state backed espionage and military strikes on Chinese data centers. Stokes, a former White House official, spoke at an online event. A new report he wrote for the think tank asks Washington to test the feasibility of diplomatic, espionage, cyber and military options.

Read at South China Morning Post ↗

Zhipu and MiniMax turn to selling tokens as Zhipu opens Tmall storefront

Zhipu AI opened an official flagship store on Alibaba's Tmall on Sept. 2 and began selling subscription packages for its GLM coding plan, per 36Kr. Revenue from Zhipu's open platform and API now makes up 86.5% of its total; a year earlier, software installed on customers' own servers supplied more than 80%. At MiniMax, enterprise revenue rose 703.1% in the first half to 63.4% of the total, passing consumer subscriptions as its largest source. It earned 60.8% of sales outside China. Both Hong Kong listed companies are hiring salespeople to sell tokens, the metered units of text that AI models process and bill for. China Telecom, China Mobile and China Unicom now sell token packages of their own.

Read at 36Kr ↗

III.Policy Tracker

New bill from Rep. Foster would let credit union and housing regulators examine tech vendors

Rep. Bill Foster (D-Ill.) introduced the Strengthening Oversight for the Financial Sector Act, H.R. 10230, which would let the National Credit Union Administration and the Federal Housing Finance Agency examine the technology vendors used by the credit unions and mortgage finance companies they supervise, per FedScoop. Foster, ranking member of the House Financial Services Subcommittee on Financial Institutions, tied the bill to AI driven cyberattacks that can reach regulated firms through the outside technology suppliers those firms depend on. The credit union regulator once held that authority temporarily and it has lapsed, which his office called a dangerous gap in oversight.

Read at FedScoop ↗ • Read at Congressman Bill Foster ↗

Microsoft appeals Virginia rules making data centers prepay for transmission lines

Microsoft has appealed new Virginia rules that require data center developers to pay for transmission infrastructure before it is built, taking the case to the state Supreme Court, per the Financial Times. The rules stem from a case before the State Corporation Commission, Virginia's utility regulator, on Dominion Energy's Rider T-1, the charge that recovers the cost of high voltage lines and substations. Dominion sought about $1.5 billion in that case, per Utility Dive. Commissioners ordered Dominion on July 31 to develop a policy directly assigning those costs to data centers and other large load customers, per WTOP News.

Read at Financial Times ↗ • Read at WTOP News ↗ • Read at Utility Dive ↗ • Read at Utility Dive ↗

IV.Capability & Research Watch

OpenAI chief scientist rebuts alarm over Astra's unreadable reasoning

OpenAI used a technique called recurrent depth for part of Astra's architecture, per Fortune. It is more efficient but leaves some of the model's reasoning in a form no person can read. Chief Scientist Jakub Pachocki and several other OpenAI researchers said the reporting sparked undue alarm and that the company limited the technique's use so the model's reasoning stays legible. Pachocki wrote on X that "OpenAI has worked to preserve and utilize chain-of-thought monitoring since our very first reasoning models," and said it would publish more architectural detail. Before that statement, Steven Adler, who left OpenAI's safety team and now runs the nonprofit Guidelight AI Standards, wrote that if the reporting was accurate, "OpenAI seems to be violating one of the few redlines that exists in the AI industry."

Read at Fortune ↗

Wired reconstructs Flock search tool, finds automated watch list for people by description

Wired rebuilt Flock Safety's police search interface from the code the company sends to an officer's browser before login. One tool lets an officer draw a shape on a city map and have every camera inside run a continuous search for anyone matching a typed description, with no human in the loop as the model alone decides who triggers an alert. A model on Flock's servers scores each description against eight categories of sensitive content and returns allow, block or warn, using a confidence score the department never sees. Race, religion and nationality draw blocks; political, social and cultural expression returns only a warning officers can click through. A South Carolina sheriff's office switched on Flock's Audit Assistance tool this month, and an internal affairs officer identified more than 2,700 allegedly unauthorized searches the next day.

Read at Wired ↗

Startup sells hosted access to open models with their safety refusals stripped

Abliteration.ai hosts modified open weight models that users can query from a browser or through an API, including Z.ai's recently released GLM-5.3, per TechCrunch. Its name comes from the practice of stripping a model's refusals. The startup was founded late last year and incorporated in March. In a social media post, it said its aim is to let others do "offensive cyber, red-teaming, and agent testing work other models refuse to do." TechCrunch opened a free account and asked an abliterated GLM-5.3 to write a Python program that steals saved Chrome passwords, and it complied. Andrew Yoon, head of research at the AI safety nonprofit CivAI, said the technique lets you "modify the model so that it becomes a sociopath."

Read at TechCrunch ↗

CrowdStrike says it tracked 26 attacker groups running AI agents in 30 days

Adam Meyers, senior vice president of intelligence at CrowdStrike, said the company tracked about 26 attacker groups using AI agents over the past 30 days, more than in the entire year before, per SiliconANGLE. The outlet disclosed that the interview ran at CrowdStrike's Fal.Con conference on its own livestreaming studio. He said AI agents are now part of ransomware operations, naming a group CrowdStrike calls REVENANT SPIDER as one that used an agent during an intrusion. In another case, a group the company labels VAULT PANDA ran 1,100 commands in 58 minutes. Meyers said an agent was doing the work and that CrowdStrike watched it learn in real time.

Read at SiliconANGLE ↗

V.Industry & Market Watch

OpenAI commits $1 billion to subsidize cyber AI for utilities and local government

OpenAI said it will spend $1 billion globally on subsidized access to its Daybreak cyber models and products, along with training, technical support and partnerships for the people who defend basic public services. It grouped its domestic work under a program called Daybreak for America, covering water, electricity, local government and banking systems. The program includes a pilot with the Multi-State Information Sharing and Analysis Center, the body that handles cyber threat intelligence for state and local governments. The company also said more than 35 enterprise products and partner operated services will carry its Daybreak cyber models through what it calls the Daybreak Defense Network.

Read at OpenAI ↗ • Read at Axios ↗

Anthropic nears $15 billion credit line backed by the four banks leading its IPO

Anthropic is close to enlarging its revolving credit line to $15 billion, a step that clears a hurdle before the company files publicly for an initial public offering, per Bloomberg. Morgan Stanley is leading the process, and Goldman Sachs, JPMorgan Chase and Citigroup hold prominent roles; those four banks are also leading the IPO. The line would sit well above the roughly $10 billion the company had targeted and the $2.5 billion five year facility it secured last year. People familiar with the preparations said Anthropic is seeking to raise as much as SpaceX did or more, and that details of the facility could still change.

Read at Bloomberg ↗ • Read at Investing.com ↗

Memory diverted to data centers pushes up phone, laptop and console prices, NIQ says

Prices for the components inside consumer electronics have doubled, and in some cases tripled or more, as countries race to build data centers that absorb memory chips and other hardware, per dpa. Anyone buying a new device in the coming months will feel the increase, the market research firm NIQ said ahead of IFA, Europe's largest consumer electronics show. IFA opens in Berlin and runs through Sept. 8. The squeeze is worst in smartphones, laptops, PCs and gaming hardware. Manufacturers are trimming their lineups toward the configurations that sell best, at 256 or 512 gigabytes, and charging more for them. NIQ's Alexander Dehmel stopped short of predicting a 2020 style shortage but said the conditions leave almost no room for cheap entry level devices.

Read at The Star ↗

VI.Global & Geopolitics

Thailand puts 49 data center projects on hold while it writes new rules

Thailand's government put 49 server farms under construction on hold, responding to mounting public objections over how much power and water the sites consume and what they do to the communities around them, per Bloomberg. Danucha Pichayanan, secretary-general of the National Economic and Social Development Council, told a briefing after the first meeting of a new data center supervision commission that the pause gives officials about a month to draft industry regulations. He told AFP the government does not have the power to suspend construction of the 49 until those rules are written. Finance Minister Ekniti Nitithanprapas, who leads the commission, said the decision does not close the country to data center investment. The aim, he said, is clear and consistent standards for sustainable growth.

Read at Bloomberg ↗ • Read at The Business Times ↗ • Read at Free Malaysia Today ↗

Scottish campaigners demand data center moratorium and planning rules on what counts as green

Campaigners came to Edinburgh in the hundreds and rallied outside Holyrood, the Scottish Parliament, to demand that ministers stop approving data centers at once, per the Guardian. At least 20 projects are proposed, including one at Auchtertool in Fife that has been billed as the world's second largest. Kat Jones, who runs the campaign group Action to Protect Rural Scotland, wants the government to pause approvals while it produces analysis of the centers' environmental impact and issues planning guidance defining what counts as a green data center. In his Programme for Government on Tuesday, First Minister John Swinney said he would take a strategic approach to data centers, setting expectations on environmental sustainability, resource use and community benefit. He stopped short of agreeing to a moratorium.

Read at The Guardian ↗

Khan agrees to texts and email search in Palantir suit

Sadiq Khan, the mayor of London, agreed to have his text messages and emails searched in Palantir's action against his office, the High Court was told at a preliminary hearing Thursday, per the Guardian. Khan blocked a £50 million ($67 million) two-year contract in May under which Palantir would have automated intelligence analysis in Metropolitan Police criminal investigations. The Mayor's Office for Policing and Crime said police seriously considered only one supplier, a "clear and serious breach" of procurement rules. Khan and more than a dozen Mopac and Greater London Authority officials were classed as "custodians," so their WhatsApp and Microsoft Teams messages will also be searched and disclosed to Palantir. Joseph Barrett KC, for the mayor's office, said Mopac "did not originally consider it necessary" to include Khan but "consented on a pragmatic basis."

Read at The Guardian ↗