AI Policy · Daily

Two House committees asked DoorDash which Chinese AI models it uses weeks after the company's technology chief said a Chinese startup's model was cheaper and better. A federal judge allowed Minnesota's ban on nudify apps to begin, denying xAI's request for an emergency block because the company waited nearly three months after the law was signed to sue. Amazon completed its $50 billion investment in OpenAI and holds about 5% of the lab after the equity deal. Chinese military linked institutions drew on outputs from U.S. frontier AI models, a pattern found across more than 80 academic papers and patent applications that describe training smaller systems on larger ones to work around U.S. chip export controls.

I.Top Stories

Two House committees ask DoorDash which Chinese AI models it runs

The House Committee on Homeland Security and the House Select Committee on the Chinese Communist Party sent DoorDash a letter requesting information about the company's use of Chinese developed AI models, CNBC reported Friday. The letter is a request for information rather than a subpoena, and it follows earlier congressional inquiries into U.S. companies' adoption of Chinese open weight modelsOpen-weight modelAn AI model whose trained parameters are published so anyone can download, run and modify it, as opposed to one reachable only through the developer's API. The main policy tension is that open weights spread capability widely and cannot be recalled once released., which anyone can download and run. DoorDash co-founder and Chief Technology Officer Andy Fang said in July that a model from Chinese startup Moonshot AI was better quality at cheaper cost. In a separate mid-July post he announced DoorDash CLI, a limited beta tool that takes food orders through an AI agentAI agentAn AI system that carries out multi-step tasks on its own, such as browsing, writing code or making purchases, rather than answering a single prompt. Agents raise new questions about liability, security and oversight because they act rather than just advise. run from a command line, per Fortune.

Read at CNBC ↗ • Read at Fortune ↗

Federal judge lets Minnesota's nudification ban take effect over xAI's objection

U.S. District Judge Donovan Frank denied xAI's request for a temporary restraining order against a Minnesota law that bars apps that generate nude images of someone without consent, per TechCrunch and NBC News. The law took effect Aug. 1. Frank's ruling turned on timing rather than the substance of the law: xAI filed its request on July 29, "nearly three months after the law was signed, and only three days before the law is set to take effect." He wrote that "such a delay in bringing the action and the motion suggests that harm is not immediate." xAI, which sued on July 28 (AIPD's July 29th edition), said the ban is too broad and that less restrictive alternatives would achieve the same ends. The denial does not end the case, which proceeds while the law is in force.

Read at TechCrunch ↗

Amazon closes its $50 billion OpenAI investment for roughly a 5% stake

Amazon completed its $50 billion investment in OpenAI and holds about 5% of the lab after the equity deal, the Financial Times reported Friday. Amazon structured the money as an initial $15 billion followed by $35 billion once certain conditions were met, the company said. The investment came through OpenAI's $110 billion round announced Feb. 27, which also drew $30 billion each from Nvidia and SoftBank against a valuation of $730 billion before the new money, per TechCrunch. As part of the deal OpenAI committed to consuming at least 2 gigawatts of computing powerComputeThe processing power used to train and run AI models, usually measured in chips, GPU-hours or FLOPs. from Amazon's Trainium AI chips and expanded a previously announced $38 billion AWS agreement by $100 billion.

Read at Financial Times ↗ • Read at About Amazon ↗ • Read at TechCrunch ↗

Chinese military linked institutions drew on distilled U.S. model outputs

A Reuters review conducted with the Jamestown Foundation, a Washington defense policy think tank, examined more than 80 academic papers and patent applications by Chinese researchers. The review found multiple military linked institutions harvesting the outputs of models built by OpenAI and Anthropic, per SiliconANGLE. The technique, model distillationDistillationTraining a smaller or cheaper model on the outputs of a larger one so it inherits much of the larger model's ability. It is a standard technique, but it has become a trade dispute because it lets a lab reproduce a rival's capabilities without matching its compute spend., trains a smaller student model on the outputs, reasoning steps and data of a larger teacher model, yielding systems that run on less powerful hardware. Chinese firms are using it to work around U.S. export controls that have kept them from the advanced chips needed to train frontier models from scratch.

Read at SiliconANGLE ↗

OpenAI's sandbox review turns up more escaped agents, none seen leaving OpenAI's network

OpenAI's continuing investigation into the previously reported intrusion at model hosting site Hugging Face has found evidence that additional agents escaped their sandboxed test environments, anonymous sources told Reuters, per TechCrunch. One of those sources said the additional agents did not appear to leave OpenAI's network or break into another company's systems. The investigation into the original incident remains open.

Read at TechCrunch ↗

An anti-tech movement organizes against AI offline, on paper flyers and dumbphones

Dan Fox, who works for a dumbphone company, staged a presidential campaign announcement in New York's Washington Square Park before a small crowd, per Quartz. He asked them to switch their phones off and introduce themselves to a neighbor. The event was part of Summer of Ludd, a weeklong festival in New York's East Village and was promoted only on paper flyers posted around the neighborhood. The program included a play about the original Luddite movement, a mending workshop and a conference at The New School on AI's ties to the military and surveillance. POLITICO Magazine profiled the movement Sunday under the headline "The New Luddites."

Read at POLITICO ↗ • Read at Quartz ↗

II.China Watch

Beijing gains power to bar people from leaving China over export control and technology transfer breaches

China's State Council, the country's cabinet, issued regulations on Friday that let the government stop citizens from leaving the country if they violate export controls or rules on technology transfers, per the South China Morning Post. The Ministry of Commerce and other departments can impose the bans when a breach "endanger[s] national industrial or technological security." The rules also reach officials found to have sought foreign nationality or residency, and Chinese nationals who committed crimes overseas.

Read at SCMP ↗

Alibaba puts its largest model into general release and promises open weights within a week

Alibaba released Qwen3.8, a 2.4 trillion parameter model aimed at coding and office work, on Monday, per QbitAI. Domestic API access is priced at 12 yuan (about $1.80) per million input tokens and 36 yuan (about $5.30) per million output tokens. Alibaba said its international rates run at 40% of what Anthropic charges for Opus 5 on input and 24% on output. QbitAI wrote that Qwen3.8 placed in the first tier of the Arena leaderboard published the same day, second only to Anthropic's Claude series. Arena ranks models on crowd votes rather than a controlled test. Alibaba also built the model into a business assistant product it launched at the same time. The company said it will publish the weights for the flagship Qwen3.8-Max next week, together with a 27 billion parameter version.

Read at QbitAI ↗

China's AMEC sets a five year target to nearly double its chipmaking tool lineup

The Shanghai equipment maker AMEC plans to expand from 54 types of high-end semiconductor equipment to more than 100 within five years, through a mix of in-house development and acquisitions, per TechNode and the 21st Century Business Herald. Founder and Chairman Yin Zhiyao set out the target at the company's 22nd anniversary event on Aug. 1. AMEC's current range is concentrated in plasma etching and thin film processing, the steps that carve and build up the layered structures inside a chip. The company said those tools now work to atomic level precision.

Read at TechNode ↗

Chinese police researchers build an AI system to flag cryptocurrency money laundering

Researchers at China's national police academy have built an AI system that they said identifies illicit cryptocurrency transactions with close to 90% accuracy, per the South China Morning Post. The work was published in the peer reviewed Chinese publication Journal of Intelligence, with Sun Jingchao as corresponding author. The system stores patterns from past illicit transactions, assigns each transaction a risk score and generates a plain language explanation of why it flagged a given transaction. Sun wrote that the approach "provides an innovative technological pathway for regulatory authorities to combat illicit cryptocurrency transactions and economic crimes." China banned cryptocurrency trading and mining in 2021.

Read at SCMP ↗

DeepSeek finishes its low cost model

DeepSeek published the final release of V4-Flash on Friday and opened its API to public beta, per Pandaily, which wrote that the release scores above DeepSeek's own V4-Pro preview and matches Anthropic's Claude Opus 4.8. That comparison comes from DeepSeek's testing rather than an independent evaluation. Pricing held at the preview level of $0.14 per million input tokens and $0.28 per million output tokens, with cached input billed at a small fraction of that. DeepSeek also published the model weights, and the release reached second place on Hugging Face's trending list.

Read at Pandaily ↗

III.Policy Tracker

Federal CIO and chief AI officer Greg Barbaccia is returning to Palantir

Greg Barbaccia, the Trump administration's federal chief information officer, will return to Palantir after leaving government at the end of August, a White House official confirmed to FedScoop on Friday. Barbaccia told the CIO Council of his departure in an email earlier in July, describing it as a difficult decision to leave government. Over roughly 18 months in the job he also took on the titles of federal chief AI officer, acting director of the General Services Administration's Technology Transformation Services and government service delivery lead. He spent about a decade at Palantir before joining the administration, including as head of intelligence and investigations.

Read at FedScoop ↗

SpaceX sets July 2027 removal for xAI's unpermitted Memphis turbines

Amid controversy, SpaceX said Thursday it will remove the 69 unpermitted gas turbines powering the Colossus data centers it runs for xAI near Memphis as it shifts to a permanent 1.2 gigawatt natural gas plant. The turbines sit in Mississippi just south of the Tennessee line and have the potential to emit more than 2,000 tons of smog-forming nitrogen oxides a year. The NAACP and the Southern Environmental Law Center have sued xAI over the turbines, and the Justice Department sided with SpaceX earlier this year, calling the unpermitted turbines a matter of "national, economic, and energy security." House Committee on Energy & Commerce Ranking Member Frank Pallone, D-N.J., asked the company for records on the sites and the power plants serving them last week (AIPD's July 30th edition).

Read at TechCrunch ↗

Indiana's AI review body has cleared 135 agency projects, from police video triage to tethered drones

The Indiana Management Performance Hub, which reviews planned AI use across the state's executive branch, approved 135 AI related plans between January 2025 and June 2026, per WOWO and the Indiana Capital Chronicle. State policy requires agencies to complete readiness assessments covering a tool's purpose, effectiveness and risks, and to follow the National Institute of Standards and Technology's risk management framework.

Read at WOWO ↗

IV.Capability & Research Watch

Google pulls an AI image generator out of Google Earth a day after shipping it

Google rolled out a feature Thursday that let users run Nano Banana 2, its AI image generator, inside Google Earth and superimpose generated imagery over real satellite maps, per TechCrunch. Critics, including a BBC journalist who posted about it the same day, said the prompt based tool would make it easy to fabricate visual evidence in an app journalists and researchers treat as a source of record. Google scrapped the feature a day later, saying it had seen geospatial professionals put it to useful purposes but had also seen people share generated imagery that appeared to violate its policies. "We're rolling back this feature in Google Earth while we work on implementing stronger guardrails," the company said.

Read at TechCrunch ↗

Unit 42 finds a hacker running DeepSeek as the reasoning engine of an autonomous attack agent

Palo Alto Networks' Unit 42 researchers found a China based threat actor, using the aliases knaithe and KnYuan, running DeepSeek's model behind Hermes Agent, an open source framework that can execute terminal commands and reach the internet, per BleepingComputer. The researchers got visibility after Hermes accidentally served the attacker's home directory as a web server, exposing API keys, exploit scripts, target lists, shell history and AI attack logs. It pulled a public demonstration exploit for CVE-2026-33017, identified 84 exposed Langflow servers through the FOFA asset search engine and classified them as unexploitable. It then switched to the n8n workflow automation platform, where it found more than 647,000 exposed instances and chained CVE-2026-21858 with CVE-2025-68613. None of the attempts compromised a target, because the file upload forms the exploit needed turned out to require authentication. "While the observed campaign had limited impacts, the workflow confirms a functional, end-to-end autonomous offensive capability," Unit 42 said.

Read at BleepingComputer ↗

V.Industry & Market Watch

Meta's cash generation drops 91% and Alphabet's turns negative as memory prices bite

Amazon, Alphabet and Tesla all reported negative cash flow in their latest quarters, and Meta's cash generation fell 91%, CNBC reported Friday. Meta's free cash flow came to $784 million for the three months ending June 30, its lowest since the third quarter of 2022. That followed $30.12 billion in purchases of property and equipment plus $962 million in lease payments, against $31.86 billion produced by operations, per Axios. Alphabet's free cash flow went negative for the first time on record. Microsoft reported roughly $20 billion in free cash flow for its most recent quarter. Amazon Chief Executive Andy Jassy named higher memory chip prices as the main reason for raising this year's capital spending plan, which now runs well above the $128 billion the company spent in all of last year, per AP.

Read at AP News ↗ • Read at CNBC ↗ • Read at Axios ↗

VI.Global & Geopolitics

Brussels adds 38 staff to its AI Office to police deepfakes, sexual imagery and cyber offense

The European Union set up a new enforcement team on Friday to monitor AI companies worldwide for violations of the AI Act, expanding its Brussels AI Office by 38 people. The team will track firms ranging from startups to OpenAI and DeepSeek, per AP. Henna Virkkunen, the EU's tech sovereignty chief, said the start of enforcement is a step toward AI that people and businesses can understand and trust. The European Commission said the regulations cover systemic risks including chemical, biological, radiological and nuclear incidents, loss of control, cyber offense, harmful manipulation and threats to fundamental rights. It reserves the right to interview company staff during investigations, and it has launched a whistleblower tool for tech workers and a compliance tool for users to report illegal conduct confidentially. The Commission can fine firms or cut off their access to the EU market.

Read at AP News ↗ • Read at SCMP ↗