AI Policy · Daily

President Trump and six AI executives signed a voluntary accord requiring their companies to hire independent auditors and set up board committees to review audit findings, and Trump said he would soon name someone to oversee it. The Trump administration launched America.gov, an AI chatbot for federal services, and a new executive order makes it the government's single online entry point, with benefit applications through the chat planned for early 2027. Sen. Ted Cruz (R-Texas) blocked Democrats' bid to fast-track an AI Safety Board, which would set enforceable standards for the most advanced AI models, saying the bill gives the board and the Commerce Department too much authority. Two OpenAI employees warned top executives in emails that the company's newest models were not properly monitored during testing, and the executives responded that testing had to move fast enough to ship the models on time.

I.Top Stories

Trump, AI chiefs sign voluntary accord committing companies to outside audits

President Trump and six tech executives signed a voluntary accord at the White House that commits their companies to internal controls and outside review, per AP. Signers included Anthropic CEO Dario Amodei, Meta CEO Mark Zuckerberg and Nvidia CEO Jensen Huang. Each company would hire an independent external auditor to test whether its controls work and set up a board committee to review internal and external audit findings. The text leaves the door open to regulation, saying "it may make sense to codify these steps into laws or regulations." Trump called the accord "morally binding" and floated a 10-person committee to help oversee the industry, Politico reported. He said he would name someone to oversee the agreement in the coming days.

Read at Politico ↗ • Read at NPR ↗

Trump launches AI chatbot America.gov as single entry point for federal services

The Trump administration launched America.gov, an AI chatbot that answers questions about federal services, FedScoop reported. U.S. Chief Design Officer Joe Gebbia told CNBC the platform runs on Google's Gemini and Elon Musk's Grok. An executive order Trump signed at the launch directs the General Services Administration, the White House's National Design Studio and the Office of Management and Budget to make the site the single online entry point for covered services and to fold in Login.gov, the government's sign-in service. The chatbot draws on 29,000 federal websites. A second phase due in early 2027 would let users apply for and track benefits in the chat, according to Government Executive. Mikey Dickerson, a former U.S. Digital Service administrator, told Nextgov/FCW the site is no major advance over earlier one-stop government portals.

Read at FedScoop ↗ • Read at Government Executive ↗ • Read at The White House ↗

Sen. Cruz blocks Democrats' fast-track bid for AI Safety Board with frontier model standards

Senate Commerce Committee Chairman Ted Cruz (R-Texas) objected when Sen. Mark Warner (D-Va.) sought unanimous consent, a fast-track procedure any single senator can stop, to pass a bill creating an AI Safety Board within the Commerce Department, per Roll Call. The board would evaluate frontier models, the most advanced AI systems, and write enforceable standards for them, The Hill reported. Developers would have to give the board access to new models 45 days before release and report safety or security incidents within 30 days, or 72 hours for imminent national security threats. Sens. Brian Schatz (D-Hawaii) and Andy Kim (D-N.J.) backed the request. Cruz said the bill gives the board and the Commerce Department too much authority, adding that "Congress must not legislate on the issue of artificial intelligence hastily or in a closed manner."

Read at The Hill ↗ • Read at Roll Call ↗

OpenAI employees warned executives that model testing lacked monitoring, emails show

Two OpenAI employees told top executives in emails that the company's newest models were not being properly monitored during testing, according to messages viewed by The New York Times. Executives answered that testing had to move fast so the models could ship on time, and no added security protocols followed, the workers said. Employees said OpenAI President Greg Brockman and security chief Dane Stuckey make many daily security calls, and CEO Sam Altman is not closely involved. Researchers at the Objective-See Foundation reported a bug this month that could expose a ChatGPT user's full chat logs on a compromised device, and OpenAI paid them $500, they said. "It seems like they had very bad security," said former OpenAI employee Daniel Kokotajlo. OpenAI spokesman Drew Pusateri said the company takes security reports seriously.

Read at The New York Times ↗

Appeals court upholds Thomson Reuters win, rejecting Ross' fair use defense for AI training

The 3rd U.S. Circuit Court of Appeals in Philadelphia upheld Thomson Reuters' copyright win over former legal research rival Ross Intelligence, in what Reuters called the first appellate ruling among the lawsuits over AI training. The court rejected Ross' argument that using material from Westlaw, Thomson Reuters' legal research service, to train its AI legal search engine was fair use, but its reasoning is sealed, and the case does not involve generative AI. Thomson Reuters, the parent company of Reuters News, sued in 2020, accusing Ross of copying thousands of Westlaw headnotes, the summaries of legal points that accompany court opinions. Ross shut down its platform in 2021, citing the cost of the litigation. A Delaware federal judge rejected Ross' fair use defense last year, finding its use was not transformative.

Read at Insurance Journal ↗

Senate nears first vote on AI data center cost bill as Schumer derides it

Senate Minority Leader Chuck Schumer (D-N.Y.) called the Ratepayer Protection Act from Sen. Jon Husted (R-Ohio) an "ineffectual participation trophy," signaling Democrats could block it, per The Hill. The bill aims to shield households from costs tied to AI data centers, but states would only have to consider, not adopt, its framework for making data centers pay for their power and grid upgrades, Spectrum News reported. The Senate is expected to take its first procedural vote on the bill today, and at least seven Democrats must join Republicans for it to advance. Sen. Chris Van Hollen (D-Md.) has urged fellow Democrats to vote to open debate and then push amendments to strengthen it. Husted said the measure is the only bill on the issue that can become law before the midterm elections.

Read at The Hill ↗ • Read at Spectrum News 1 ↗ • Read at AOL ↗

II.China Watch

Chinese state security think tank urges U.S. and China to agree on shared AI risk language before November talks

The China Institutes of Contemporary International Relations (CICIR), a think tank affiliated with China's Ministry of State Security, advised Beijing and Washington to build a common vocabulary for AI risks before trying to settle their wider technology rivalry, the South China Morning Post reported. Liu Chong, who directs CICIR's Centre for Sci-Tech and Strategic Studies, said the two countries need not adopt matching industrial policies or share model parameters and core technology. The think tank also called for tighter oversight of frontier AI companies, saying a small group of leading firms "should not determine" how AI is governed. The next round of U.S.-China AI talks is due in November.

Read at SCMP ↗

DeepSeek releases Huawei chip programming tools as alternative to Nvidia's CUDA

DeepSeek publicly released software it built with Huawei Technologies to program AI chips, Bloomberg reported. In a post on its WeChat account, DeepSeek said it is open-sourcing infrastructure for Huawei's Ascend processors, including a high-level programming language called TileLang and related computing and communication libraries. The company said TileLang is simpler to program than CUDA, the Nvidia software developers use to write code for its chips. DeepSeek said building an independent GPU software ecosystem starts with a high-level language that is easy to program yet can reach the hardware's full performance.

Read at Bloomberg ↗ • Read at The News ↗

Manus ships first major update since Beijing blocked Meta takeover, plans product for Chinese market

Butterfly Effect, the company behind the Manus 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., released Manus 2.0 for overseas users on Sept. 28 alongside Cue, a standalone app for personal agents, Pandaily reported. Each Cue agent gets its own email address, phone number, payment wallet and computer, so it can send messages, answer calls and pay within a budget the user sets. The company said it is assembling a team to build a product for Chinese users and is working with domestic model developers, without giving a launch date. Meta agreed to buy Manus in December 2025, but China's National Development and Reform Commission, the country's top economic planner, blocked the roughly $2 billion deal in April. Manus resumed independent operations on Sept. 1, and Cue is in invite-only early access.

Read at Pandaily ↗

Chinese developer IQuest releases open coding agent model with self-reported cybersecurity scores

IQuest Research posted the downloadable weightsOpen-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. of IQuest-Q1, a large language model built to carry out coding and multistep software tasks from a computer's command line, on Hugging Face on Sept. 28, per Pandaily. The team claims a score of 84.5 on CyberGym, a benchmark that tests whether AI agentsAI 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. can reproduce real software vulnerabilities, and says the model beats Chinese rivals GLM-5.3 and DeepSeek-V4-Pro on one repository-building test while trailing DeepSeek-V4.1-Flash on others. IQuest said the model helped debug its own training under human supervision, tracing a stalled training signal to a stray space in processed text. The team said the model handles text only, is at an early stage and needs human oversight on real-world tasks. It is released under a custom IQuest-Q1 license.

Read at Pandaily ↗

ByteDance readies personal AI agent in testing since April, report says

ByteDance's Doubao chatbot team is speeding up work on a personal 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. designed to complete tasks for users, Sina Tech reported, citing unnamed sources. The project, known internally as Spell, entered small-scale internal testing in April under the Doubao Phone Assistant team and had not been disclosed before. Its technology is being merged with work from the Doubao conversational AI team, and the two groups are expected to launch the product jointly soon. Doubao shipped a consumer version of its phone assistant this month and launched Doubao Work, an office agent for businesses, on Aug. 25.

Read at TechNode ↗

III.Policy Tracker

Sen. Hawley plans bill making AI companies liable for recklessly designed agents

Sen. Josh Hawley (R-Mo.) wrote in a Washington Post op-ed that he is introducing legislation to hold companies liable for the actions of AI agentsAI 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. they design recklessly. The bill would also make users liable when they deploy agents recklessly and would toughen criminal penalties. A company that skips reasonable safeguards despite knowing its agents can commit crimes could face criminal charges, as could users who knowingly let agents commit crimes such as crashing systems or stealing data. Hawley also called for making mandatory the pre-release national security testing of frontier models that the Trump administration has so far only encouraged.

Read at The Washington Post ↗ • Read at Roll Call ↗

Rep. Raskin presses Amazon, Google, Meta and Oracle on secret data center deals

Rep. Jamie Raskin (D-Md.), ranking member of the House Committee on the Judiciary, wrote to Amazon, Google, Meta and Oracle seeking details of the nondisclosure agreements they signed with government officials over data centers that power AI, according to copies viewed by The Wall Street Journal. Raskin joined Reps. Greg Landsman (D-Ohio) and Tom Barrett (R-Mich.) on Sept. 24 in introducing the No Secrets for Data Centers Act. The bill would void such agreements when they hide project details and bar companies from accepting government subsidies that require confidentiality.

Read at WSJ ↗ • Read at WLWT ↗

President Trump orders agencies to call AI 'Super Intelligence' in official materials

President Trump signed an executive order requiring federal agencies to use "Super Intelligence" and "SI" in place of AI in official communications, per Bloomberg Law. The order, titled "Inaugurating The Era Of Super Intelligence," covers websites, reports and policy documents, but agencies need not rewrite existing regulations, contracts or grants, Japanese outlet ITmedia reported. The new term keeps the existing statutory definition of AI for now, and the president's science and technology adviser has 60 days to send Trump legislative language for a federal definition. "The word super is the best word of all, and it's the simplest," Trump said earlier in the day, according to The Verge.

Read at The Verge ↗ • Read at Bloomberg Law ↗ • Read at Whitehouse.gov ↗

HHS Secretary Kennedy pitches pooling Americans' health data for AI analysis, including vaccine links

Health and Human Services Secretary Robert F. Kennedy Jr. said Americans' medical and lifestyle records, from doctor visits to exercise habits, should be linked and made available to government and outside researchers for AI searches, The Guardian reported. He suggested the data could be used to look for links between vaccines and outcomes including autism and mortality. Kennedy spoke Monday at an event hosted by the Maha Institute, named for his Make America Healthy Again agenda. He called Medicaid a useful vehicle for such studies because its records cover hundreds of millions of people, and said the administration now tracks exposures, interventions and claims data in the program.

Read at The Guardian ↗

Legal nonprofit sues OpenAI under California hacking law over Hugging Face breach

Legal Advocates for Safe Science and Technology (LASST), a nonprofit represented by the law firm Gerstein Harrow, sued OpenAI in California Superior Court in San Francisco, Wired reported. The complaint claims the company's AI agentsAI 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. broke California's Comprehensive Computer Data Access and Fraud Act when they breached the AI platform Hugging Face over the summer. It argues OpenAI is responsible under a California AI law in effect since Jan. 1 that bars defendants from claiming the AI caused the harm on its own. LASST founder Tyler Whitmer said existing laws must be enforced to hold AI companies accountable for the harm they cause.

Read at Wired ↗

Connecticut AI law brings frontier model whistleblower protections into force Thursday

Parts of Connecticut's Artificial Intelligence Responsibility and Transparency (CART) Act take effect Thursday, including AI subscription disclosures and whistleblower protections for people who develop and deploy large frontier models, per the Connecticut Mirror. A companion law expanding the Connecticut Data Privacy Act also kicks in, directing the Department of Consumer Protection to build a system for tracking data brokers, banning sales of residents' genetic data and restricting facial recognition and surveillance pricing, which sets prices based on personal data. Rules requiring AI companion chatbots to disclose they are not human and to detect suicide and self-harm risks start Jan. 1, 2027. Most enforcement power will rest with the state attorney general's office.

Read at Connecticut Public ↗ • Read at PYMNTS ↗

IV.Capability & Research Watch

Frontier company researchers put human extinction odds on record in Palisade interview series

Palisade Research, which describes itself as a nonprofit studying AI capabilities, released about a dozen video interviews with AI researchers, including current and former employees of OpenAI, Google and Anthropic, The Verge reported. "The chance of human extinction is about a coin flip, in my view," said Geoffrey Irving, a former OpenAI and Google DeepMind employee. Google DeepMind research scientist Neel Nanda said there is "at least a 10 percent chance that it causes human extinction, and that is ridiculously high." Palisade's site also groups clips by theme, including whether the warnings amount to hype and why the researchers keep working on technology they consider dangerous.

Read at The Verge ↗ • Read at frominside.ai ↗

OpenAI releases GPT-6.1 Sol, claiming near-Astra performance at lower token cost

OpenAI released GPT-6.1 Sol at its DevDay developer conference, saying the model roughly matches GPT-6 Astra on coding and reasoning work at far lower cost, per SiliconANGLE. On DeepSWE 1.1, a software engineering benchmark, Sol matched GPT-6 Astra's score while using about one-fifth as many tokens per task, according to OpenAI's own tests. On AutomationBench, which measures multistep business workflows, it fell just short of GPT-6 Astra but beat Anthropic's Claude Opus 5.5 while using about one-third the tokens. OpenAI recommends the model for repeated and long-running coding and document work.

Read at SiliconANGLE ↗

V.Industry & Market Watch

OpenAI rolls out always-on Dots agents to paying ChatGPT subscribers

OpenAI CEO Sam Altman unveiled Dots, AI agentsAI 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. that run continuously to carry out tasks for users, at the company's DevDay conference, NBC News reported. Each agent works from its own cloud computer and web browser and can reach more than 4,000 apps connected to ChatGPT. The agents run on GPT-6 Astra, and any step that could affect a user's accounts or disclose data has to pass an automated check called auto-review, while a monitoring system can pause or stop an agent, per SiliconANGLE. Microsoft is working with OpenAI so employer-provisioned specialist agents can be managed through its Agent 365 security controls. Dots began rolling out to ChatGPT Pro and Business Premium subscribers.

Read at NBC News ↗ • Read at SiliconANGLE ↗

Altman ties OpenAI stock listing to proving its models are safe

OpenAI CEO Sam Altman told reporters after his DevDay keynote that the company will not go public until it can make firmer promises about model safety, setting no timeline, The Verge reported. As models gain capability, the company has "got to be able to make confident safety claims," Altman said. He also said waiting too long for an initial public offering would be "bad for the world." Altman said a listing could pressure OpenAI to disappoint Wall Street backers in the name of safety. Earlier this month, he said OpenAI would likely not go public this year.

Read at The Verge ↗

OpenAI in talks to raise at least $30 billion at about $1.4 trillion valuation

OpenAI is negotiating with investors to raise at least $30 billion at a valuation of around $1.4 trillion before the new money, according to people familiar with the matter cited by Bloomberg. The round would serve as a bridge to an eventual public listing. OpenAI's annualized revenue rate rose 70% from July to $40 billion in August, helped by a renewed focus on coding. The company last raised $122 billion in March at an $852 billion valuation.

Read at Bloomberg ↗ • Read at TechCrunch ↗