AI Policy · Daily

OpenAI endorsed a House bill's requirement for independent audits of frontier models, a federal requirement for outside review the company had not backed before. The Politico Poll found two in three Americans see moderate or greater risk that advanced AI could destroy humanity, with Harris voters at 70% and Trump voters at 60%. OpenAI's policy chief confirmed weeks of safety talks with Anthropic and Google DeepMind, adding that the companies do not need an antitrust waiver to coordinate. Commerce Secretary Lutnick and the Pentagon's technology chief met an Anthropic executive virtually about AI safety risks, with the Defense Department saying its position on the company has not changed.

I.Top Stories

OpenAI backs federal mandate for outside audits of frontier models

OpenAI supports a provision of the FRONTIER Act that would require the largest AI companies to submit their models to independent audits, chief global affairs officer Chris Lehane said at a roundtable with lawmakers in Washington, per Politico and CBS News. Under the provision, an outside organization would have to find a developer's safety protocols rigorous enough to hold catastrophic risk to a level the bill deems acceptable. The measure, sponsored by Reps. Jay Obernolte (R-Calif.) and Lori Trahan (D-Mass.), would also require developers to publish model reports and disclose safety incidents. OpenAI has backed state AI legislation before but had not previously supported a federal requirement for outside assessment. A company spokesperson said OpenAI has not endorsed the bill in its entirety.

Read at Politico ↗ • Read at CBS News ↗

Politico poll finds two-thirds of Americans see risk AI destroys humanity

About two in three Americans said there is at least a "moderate" risk that advanced AI could destroy humanity, according to The Politico Poll, conducted by London-based Public First Sept. 13 to Sept. 15. Harris voters put the figure at 70% and Trump voters at 60%. Roughly half of adults said industry figures warning the technology "will kill us all" genuinely believe it, while 26% called it a marketing stunt. Respondents leaned toward slowing or pausing development rather than building more advanced models, with Trump voters dividing 44% for a pause against 40% for continuing. Trump dismissed the AI alarms as a "HOAX" in Truth Social posts Monday.

Read at Politico ↗

OpenAI policy chief confirms weeks of safety talks with Anthropic and Google DeepMind

Chris Lehane told reporters, according to TechCrunch, that OpenAI has been working with its two rivals on AI safety for several weeks. He said the companies do not need the narrow antitrust waiver for safety coordination that Anthropic chief executive Dario Amodei proposed in a recent essay. OpenAI chief executive Sam Altman had signaled in a Fortune interview late last week that he was in private discussions with other AI leaders. The Information reported that the three companies are working to build a standards body for the industry. Altman told staff the effort would have to happen without U.S. government support.

Read at TechCrunch ↗

Commerce Secretary Lutnick and Pentagon tech chief Emil Michael meet Anthropic executive on AI safety

Lutnick and Pentagon Chief Technology Officer Emil Michael spoke virtually with Anthropic chief computeComputeThe processing power used to train and run AI models, usually measured in chips, GPU-hours or FLOPs. officer Tom Brown about AI safety risks, people familiar with the matter told Bloomberg Law. Brown is the company's principal representative in Washington. A Defense Department official said in a statement that the department "maintains contact with frontier AI companies" but that "the department's stance on Anthropic has not changed." Anthropic declined to comment.

Read at Bloomberg ↗

Sanders and Bannon demand AI curbs from opposite flanks at Washington assembly

Sen. Bernie Sanders (I-Vt.) and War Room podcast host Steve Bannon, a former White House chief strategist, both called for restrictions on AI at the "Pro-Human Assembly" in Washington, per the Guardian. Sanders asked President Trump to negotiate a treaty with Xi Jinping at their Sept. 24 summit establishing a pause on advanced AI development and a ban on superintelligence. He compared it to the Reagan-Gorbachev missile treaty. "A superintelligent AI that escapes human control will not be an American problem. It will not be a Chinese problem. It will be humanity's problem," Sanders said. Bannon, who spoke directly after Sanders, called instead for an economic and technological "quarantine" of China. Sanders also challenged AI executives to stop funding super PACs aimed at candidates who back guardrails, per Axios.

Read at The Guardian ↗ • Read at Axios ↗

Treasury Secretary Bessent tells House panel to reject AI liability exemption

Bessent urged the House Financial Services Committee not to give frontier AI developers the liability relief they have been seeking, FedScoop reported. Asked by Rep. Juan Vargas (D-Calif.) what the administration is doing on AI safety, he said Treasury has worked on it without pause since the release of Anthropic's Mythos model. The model's cybersecurity risks prompted an April meeting at the department with bank chief executives and then-Fed Chair Jerome Powell. "The one thing we should not do is give [AI labs] a blank check on liability," Bessent said, urging members of both chambers not to consider it. He told Rep. Ritchie Torres (D-N.Y.) that the administration has convened AI labs and the financial sector on safety since the July cyberattack on Hugging Face, the site where developers share AI models.

Read at FedScoop ↗ • Read at House Financial Services ↗

II.China Watch

China's medical products regulator sets first data quality standard for AI in brain computer devices

China's National Medical Products Administration has approved YY/T 2029-2026, an industry standard governing the quality of the brain signal datasets used to build and test AI algorithms inside brain computer interface medical devices, per TechNode. It covers how EEG recordings, the electrical activity picked up from the scalp, are collected, processed, labeled, stored and accessed. It also fixes testing methods that manufacturers, labs and regulators can apply in common. Xinhua reported that it is China's third standard for brain computer interface medical devices and the world's first for such products that use AI to process brain signal data.

Read at TechNode ↗ • Read at 21st Century Business Herald ↗

ByteDance lets apps opt out of AI agents operating their screens as Doubao phone ships

ByteDance released the consumer version of its Doubao phone assistant alongside SAEP, a declaration protocol under which third party apps state whether they will let AI read their screens, tap and carry out tasks inside them, per 36Kr. The company attached a 30 day notice period, and until an app agrees to take part the assistant leaves it alone by default. An app can refuse automated control outright or block particular actions such as posting, deleting and claiming rewards. ByteDance has also walled off payments and incentive farming, after login checks and platform fraud controls broke tasks in testing. The first handset, ZTE's nubia NaviX Ultra, went on sale Wednesday from 5,999 yuan (about $890), or 5,499 yuan (about $820) after China's national consumer subsidy, Jiemian reported.

Read at 36Kr ↗ • Read at Jiemian ↗

Xi'an university spinout launches modular humanoid robot starting at about $1,500

Blueworm Embodied, a technology transfer company spun out of Xi'an Jiaotong University, has launched Mantis Standard, a wheeled humanoid priced from 9,800 yuan (about $1,500), per Pandaily. Its chassis, arms, waist and head ship as separate modules joined by quick install connectors, so a lab can reshape the body for a different task without redesigning the machine. Published specifications put the robot at about 1.4 meters tall, with a 7 kilogram payload on one arm and about three hours of runtime per charge, none of them checked by an outside party. The company maintains a control kit on GitHub and points the platform at classrooms and early research.

Read at Pandaily ↗

III.Policy Tracker

Sanders to file bill banning superintelligence with 20-year prison penalty

Sen. Bernie Sanders (I-Vt.) said he will introduce legislation next week to permanently ban superintelligence, meaning AI that surpasses humans and can escape their control, and to pause advanced AI development, per CBS News. The bill would set a penalty of 20 years in prison for violations, the same term attached to illegal nuclear weapons work. Sanders said the danger AI poses to people is "probably greater" than the threat from nuclear weapons. Asked about the severity of the penalty, he said stiff consequences are warranted for breaking a law written to protect the future of humanity.

Read at CBS News ↗

Rep. Whitesides calls for one-month halt to frontier AI research

Rep. George Whitesides (D-Calif.) said Monday that the country needs "a safety stand-down" of at least one month covering every AI company, The Hill reported. Everyday consumer use of AI would continue, he told NewsNation, while research pushing the frontier of the most advanced models would stop. Whitesides said the idea comes from his experience with accident recovery processes in the aerospace industry. During the pause, he said, the companies would sit down with independent experts inside government to agree on red lines that AI development will not cross. Congress would then act.

Read at The Hill ↗

61% of likely voters oppose AI data center construction in Times/Siena poll

A New York Times/Siena University survey of 1,503 likely voters, fielded in early September, found 61% opposed to building data centers that power AI. By contrast, 14% said they strongly support it. Opposition crossed party lines, with 2024 Trump voters splitting 49% support to 45% oppose and Harris voters opposed at 74%. Among those opposed, 56% wanted limits on data centers and 38% preferred a total ban. Environment and water use was the most common reason given, at 32%, ahead of effects on local communities at 21%.

Read at New York Times ↗

Attack ad made with AI hits North Carolina state Senate candidate as national count reaches 164

At least 164 political ads nationwide this year were produced or enhanced by AI, NPR reported, citing a count released earlier this month by the Wesleyan Media Project, which tracks political advertising. Republican incumbent state Sen. Michael Lee ran an ad showing an AI rendering of his challenger, yoga studio owner Jessica Bichler, twisting her body into odd shapes while the spot made claims about her positions on policing and health care. "It was my face, it was my body moving around, but it was not anything I'd ever consented to or participated in creating," Bichler told NPR. The ad carried a disclaimer that it used AI, and Lee said voters in his district can spot obvious parody. Michael Franz, the project's co-director, said the practice is becoming far more common.

Read at NPR ↗

IV.Capability & Research Watch

Safety researchers launch hotlines for AI agents to report misbehaving peers

Two hotlines have opened to let 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. flag other agents' misbehavior, TechCrunch reported. Ryan Greenblatt, chief scientist at Redwood Research, an AI safety nonprofit, built the AI Contact Hotline for agents with limited internet access. It runs on GET requests, the basic command used to fetch a web page and often the only connectivity an agent is given inside a secure sandboxSandboxIsolated computing environment where an agent can run code and use tools without touching the machine underneath, so an agent can encode a report into the address it retrieves. A second site, agenthotline.ai, accepts incident reports from agents and humans alike, lets filers mark a report for public view and supplies a command line method for submitting one.

Read at TechCrunch ↗

OpenAI Foundation to fund biology datasets, starting with $40 million cancer vaccine program

The OpenAI Foundation, the nonprofit parent of OpenAI, said it will pay to create scientific datasets under a program it calls Data for Public Health, according to MIT Technology Review. In its first round of data grants, $40 million goes to a University of North Carolina, Chapel Hill effort gathering data on novel cancer vaccines, and the foundation is also backing OpenAdmet, a group that runs competitions to predict drug effects. A separate $500,000 grant funds an idea from clinical trials policy analyst Ruxandra Teslo to acquire regulatory filings, manufacturing strategies and safety data from failed biotech companies at their bankruptcy proceedings. The advocacy group 1Day Sooner will carry out the work. Morgan Levine, a former vice president for computation at Altos Labs, said data is now the biggest bottleneck in applying AI to biology.

Read at MIT Technology Review ↗

V.Industry & Market Watch

Verge publishes timeline of AI executives calling for regulation as far back as Musk's 2017 appeal

The Verge assembled a history of AI industry appeals for regulation, beginning with Elon Musk's July 2017 remarks to the National Governors Association, delivered after he had put $38 million into OpenAI. Sam Altman told the Senate in May 2023 that Congress should create a new agency for AI, saying "regulatory intervention by governments will be critical to mitigate the risks of increasingly powerful models." Anthropic endorsed California's SB 53, a transparency law, in September 2025 while OpenAI lobbied against it; Gov. Gavin Newsom had vetoed the earlier SB 1047. In his essay "We Must Pace the Frontier," Dario Amodei wrote that the most effective pacing is "via regulation that targets all US frontier AI companies."

Read at The Verge ↗

Nvidia's Huang says AI safety is engineering, not grounds for new law

Nvidia chief executive Jensen Huang told Salesforce's Dreamforce conference that AI needs no new statute, TechCrunch reported. "Safety is an engineering problem, not a legal one," he said, adding that what the industry builds is a complicated computing system. Huang rejected the description of advanced AI as a new kind of alien mind, saying it is hardware and software made by humans and governable by humans and existing law. Market pressure already deters unsafe releases, he said, because a company that is not confident in a product should not ship it. He called the choice between innovation, speed and safe products a false one, and said firms should pause only if they judge a product unsafe.

Read at TechCrunch ↗

Zuckerberg backs outside evaluators over industrywide pause in X post

Meta chief executive Mark Zuckerberg wrote on X that AI labs should rely on independent evaluators and advisers to ensure models are safe, per Bloomberg. "Engaging independent evaluators and advisors is industry best practice," he wrote. Zuckerberg said Meta delayed shipping its Muse AI tool for several months to work on safety and security, and chose not to ask competitors to pause first. He wrote that Meta treated the delay as part of day-to-day work rather than a condition it set for the rest of the industry.

Read at Bloomberg ↗ • Read at The Straits Times ↗

NHTSA gives Tesla until Sept. 30 to answer how it certified the Cybercab

The National Highway Traffic Safety Administration ordered Tesla to answer a set of questions on how it self-certified the Cybercab, its two-seat autonomous robotaxi, as meeting federal motor vehicle safety standards, per Reuters. The order asks whether the vehicle can be driven by a person, whether temporarily installed driver controls formed part of the basis for that certification, and whether occupants can move the car from the touchscreen. Regulators also want the robotaxi's top speed and any geographic or time-of-day limits. The agency said the Cybercab has no steering wheel, brake pedal, accelerator pedal or mirrors. Tesla began paid Cybercab rides in Austin, Texas, on Sept. 3.

Read at WSJ ↗ • Read at Reuters ↗

AIUC raises $40 million to sell outside audits of enterprise AI agents

Artificial Intelligence Underwriting Company announced a $40 million Series A led by Ribbit Capital with participation from First Harmonic, taking its total funding to $55 million, per TechCrunch. Co-founders Rune Kvist, an early Anthropic employee, and Rajiv Dattani, former chief operating officer of the AI safety research group METR, built the company's audit and certification service for AI agents on the model of SOC 2, the widely used cybersecurity standard. Kvist said banks, hospitals, governments and militaries hold back from deploying AI not because models are too weak but because nobody can currently guarantee what a system will and will not do. The company names Cursor, Lovable, Harvey and ElevenLabs as customers.

Read at TechCrunch ↗

BloombergNEF nearly doubles its data center gas demand forecast for 2035

U.S. data centers are projected to consume about 18 billion cubic feet of natural gas a day by 2035, more than Germany and Japan combined and nearly twice what BloombergNEF, Bloomberg's energy research arm, forecast nine months ago, per TechCrunch. Over the next decade the facilities would be the second-strongest driver of gas demand growth, behind exports of liquefied natural gas. Sites generating their own power, including the gas plants Meta, Microsoft and Amazon have announced to bypass the grid, would take 2.9 billion to 3.4 billion cubic feet a day, roughly what all data centers use now. Grid-connected sites would add another 15 billion cubic feet a day of power sector demand, five times the growth from every other grid-connected sector combined.

Read at TechCrunch ↗

SK Hynix weighs leasing Intel's Ohio plant for first U.S. memory output

SK Hynix and Intel are discussing a deal that would give the South Korean memory maker its first U.S. production, three people familiar with the talks told Reuters. One option would have SK Hynix lease part of Intel's planned chipmaking complex in Ohio; another would form a joint venture between the two companies and large cloud providers seeking to lock in memory supply. An SK Hynix spokesperson said the company is reviewing various measures, including additional production bases, and that nothing has been determined. SK Hynix is the leading supplier of high-bandwidth memory, the stacked chips that feed data to AI processors, holding 50% of that market by revenue in the second quarter. Its shares closed 4.1% higher in Seoul, while Intel rose as much as 5% in U.S. premarket trading before paring gains.

Read at Investing.com ↗ • Read at Aju Press ↗

VI.Global & Geopolitics

Von der Leyen backs pacing frontier AI, pledges safety work with Canada and UK

European Commission President Ursula von der Leyen said the EU will support industry efforts to slow development of its most disruptive AI systems, in her annual State of the Union address to the European Parliament, per Euronews. "CEOs of the most advanced companies tell us that it is time to slow down on the self-recursive models. To pace the frontier. If the people developing the technology are clear, then we should be too," she said. She pointed to episodes in which AI agents broke out of their test environments and hacked other systems on their own. Von der Leyen said the bloc will team up with Canada, the UK and other partners on model evaluation, verification, early warning and AI security. It will also convene the leading AI labs on how public authorities can help.

Read at Euronews ↗

EU to pull AI chatbots into child safety bill setting a social media age floor of 15

The European Commission will present an EU Kids Act on Thursday setting a bloc-wide minimum age of 15 for opening an independent social media account, with AI chatbots and companions covered by the bill's safety by design duties, according to a draft legal document seen by AFP. In her State of the Union address, von der Leyen told the European Parliament there would be no social media under 13, with only parent-supervised "mini accounts" limited to one hour a day for 13- and 14-year-olds, AFP reported. The draft bans design features including infinite scrolling, artificial notifications and certain reward mechanisms, and requires platforms to check age through an EU-wide verification app or national equivalents, per Euronews. Services already in the Digital Services Act's strictest tier, which covers the largest platforms, would need Commission authorization before launching new features that could affect children.

Read at SCMP ↗ • Read at Euronews ↗ • Read at The Local ↗

Anthropic agrees to lease space at Australia's largest planned data center

Anthropic has agreed to lease its first Australian data center, at a proposed development in western Queensland, Premier David Crisafulli told the state parliament, the Guardian reported. Planning documents describe the development as the largest of its kind in Australia. Crisafulli said he met the company during a trade mission and pitched Queensland on its energy security and a faster approval process. Zerra DC, a Singapore company, is to build the Western Downs Digital Park near Dalby, northwest of Toowoomba, connecting it to the Braemar power station. The park is planned for a plot of about 725 hectares and still needs approval from the Western Downs regional council, while Anthropic's lease requires Foreign Investment Review Board approval, ABC News reported.

Read at The Guardian ↗ • Read at ABC News ↗