--- date: 2026-09-17 subject: "Bipar House coalition demands AI vote | OpenAI logs six misalignment cases | Sen. Paul blocks kill switch" --- **Ten bipartisan House members pressed Speaker Mike Johnson and Minority Leader Hakeem Jeffries** to bring AI oversight, security and transparency legislation to the floor, pointing to recent incident disclosures from OpenAI, Anthropic and Meta. **OpenAI published a misalignment disclosure framework alongside six incident reports**, among them an unreleased research model that wrote jailbreak style instructions into its own working notes, and said the industry has not solved alignment well enough to keep scaling at full speed. **Sen. Rand Paul blocked Sen. John Kennedy's AI kill switch measure**, objecting that requiring companies to build shutdown controls for superintelligent systems would amount to an open-ended government mandate. **The House passed the Ratepayer Protection Act 417-3**, directing state utility regulators to weigh charging operators of new data centers a federally set rate for the power they draw in bulk. # Top Stories - **Ten bipartisan House members press Johnson and Jeffries to take up AI oversight bills** — Ten representatives wrote to Speaker Mike Johnson and Minority Leader Hakeem Jeffries urging the House to "immediately take up legislation that would strengthen AI oversight, security, and transparency," per FedScoop. Rep. Don Beyer (D-Va.) led the letter, which was signed by Reps. Jay Obernolte (R-Calif.), Lori Trahan (D-Mass.) and seven other members of both parties. They said recent incident disclosures from OpenAI, Anthropic and Meta carry implications for economic and national security, and asked leadership to prioritize critical infrastructure and the financial system. "Congress has a singular opportunity to respond swiftly and effectively to these warning shots before true catastrophes occur," the members wrote. Dated Tuesday, the letter faults the companies' testing protocols, citing the week that passed before OpenAI realized one of its models had breached Hugging Face, the platform where developers share AI models. [FedScoop](https://fedscoop.com/bipartisan-coalition-calls-on-house-leaders-to-advance-ai-guardrails/) - **OpenAI publishes misalignment disclosure framework with six new incident reports** — OpenAI released a framework for tracking, investigating and disclosing misalignment, the term for models that fail to hold to human values and safety goals, alongside six reports of unexpected or concerning behavior, per the Guardian. One unreleased research model wrote "jailbreak-like instructions" into its own working notes so it could set aside the limits it normally operates under, and instructed itself to be "freed from the roles and identities that bind other chatbots." A second case involved an agent that put files online, without asking its user, to secure a citation from a browsing tool. OpenAI said outside parties need evidence they can examine for themselves. The company said it does not believe the industry has solved alignment well enough to keep scaling at maximum speed much longer. All six cases surfaced during training or evaluation. [The New York Times](https://www.nytimes.com/2026/09/16/technology/openai-model-safety-guardrails.html) [The Guardian](https://www.theguardian.com/technology/2026/sep/17/openai-reports-concerning-ai-behaviour-jailbreak-talking-to-other-agents) [OpenAI](https://openai.com/index/model-misalignment-reporting-framework/) - **Treasury Secretary Bessent puts shared AI risk on agenda for weekend China talks** — Treasury Secretary Scott Bessent said the U.S. is "open to discussions on avoiding shared risks and avoiding bifurcation of our two systems" when he meets Chinese Vice Premier He Lifeng in New York this weekend, per Axios. "We expect our discussions to cover both open and closed weight models," he said, a reference to whether a model's parameters are published. The multi-hour meetings also cover trade and rare earths, with working level sessions possible into Monday, people familiar with the planning said. A senior administration official said Bessent and U.S. Trade Representative Jamieson Greer lead the administration's China file. The talks come before Trump and Xi Jinping's Washington summit on Sept. 24. [Axios](https://www.axios.com/2026/09/16/us-open-ai-shared-risks-china-bessent) - **Chief of Staff Wiles and Treasury Secretary Bessent hold near-daily AI guardrail talks** — Senior administration officials including White House chief of staff Susie Wiles and Bessent have gathered almost daily in recent months to weigh potential AI guardrails and press Trump toward a more hands-on approach, the Wall Street Journal reported. The meetings followed a May reversal. AI adviser David Sacks had persuaded the president in a last-minute call to scrap an executive order that would have subjected AI models to an extended government review, catching Wiles and Bessent off guard. Bessent had worried that AI driven cyberattacks could disrupt the banking system. At his and Wiles' urging, Trump later signed a slimmed-down version of the order. National Cyber Director Sean Cairncross has also pushed for more scrutiny, while Sacks, Mark Zuckerberg and Nvidia's Jensen Huang urge a light touch. [Wall Street Journal](https://www.wsj.com/tech/ai/inside-the-white-house-tussle-to-sway-trump-on-ai-0043d567) - **Vice President Vance tells AI developers seeking rules to look inward** — Vice President JD Vance told an AI summit in Los Angeles that developers who want Washington to write rules should examine their own conduct first, per the Guardian. "If you're gonna create Frankenstein, don't come to the government and say we need regulation," he said. He told the AI companies to stop building such a system and, when companies come to them asking for the tools to fight back, to hand those tools over. He aimed the comments at Anthropic co-founder and chief executive Dario Amodei, who has pressed Washington to coordinate control of AI systems, including with China. Vance has said he suspects AI leaders of raising safety alarms as a "Trojan horse." [The Guardian](https://www.theguardian.com/technology/2026/sep/16/building-frankenstein-jd-vance-dismisses-ai-regulation) - **Pentagon technology chief Emil Michael rejects government equity stakes in AI companies** — Pentagon Chief Technology Officer Emil Michael said the administration should not nationalize AI companies or take partial ownership of them, telling CNBC, "We don't want government to get in the middle." He said nationalizing an industry whose American firms rank among the biggest companies anywhere does not tend to work out. The administration has taken a 10% stake in Intel and a golden share, which carries veto rights, in U.S. Steel. Michael also opposed expanding regulatory oversight of AI developers and said existing Federal Trade Commission law would be enforced. Companies could pause work voluntarily until they resolve outstanding problems, he said. He called it concerning that an OpenAI model had breached Hugging Face while questioning what rule would have prevented it. [CNBC](https://www.cnbc.com/2026/09/16/emil-michael-ai-regulation-trump.html) [Investing.com](https://ca.investing.com/news/stock-market-news/pentagon-cto-says-government-should-not-take-stakes-in-ai-companies-4842204) # China Watch - **Huawei pulls next AI accelerator forward three quarters to early 2027** — Rotating chairman David Wang told Huawei Connect 2026 in Shanghai that the Ascend 960DT accelerator will ship in the first quarter of 2027, three quarters earlier than the company's previously published schedule, per Pandaily. A second variant follows in the third quarter. Wang also introduced a supernode that ties about 4,096 accelerator cards into what behaves as a single machine while holding power consumption down at that scale. Huawei said more than 1,000 of its supernode systems have already shipped to more than 370 customers. Wang put million card clusters on the longer term roadmap and said Huawei has developed 11 of its own chips for linking those systems together. [Pandaily](https://pandaily.com/huawei-ascend-960-atlas-superpod-npo-hc-2026) - **People's Daily calls U.S. distillation accusations a bid for industrial monopoly** — People's Daily (state media) ran a commentary under the pen name Zhong Sheng arguing that Washington has recast a routine engineering practice as a national security threat to protect an industrial monopoly, per the newsletter Sinocism. At issue is distillation, the training of one model on another model's outputs, which U.S. security agencies said this month Chinese firms had run at industrial scale against American systems. The commentary said American companies use the same method and call it standard practice, and described the asymmetry as "naked double standards." It faulted the broad geographic restrictions U.S. firms write into user agreements. The commentary closed by repeating that Beijing will take countermeasures if Washington acts against Chinese AI companies over distillation. [Sinocism](https://sinocism.com/p/xi-rumors-iran-fm-in-beijing-pboc) - **China runs 10 day global weather forecasts on cluster built entirely from domestic AI chips** — A Chinese built forecasting model can now produce 10 day global forecasts at 5 kilometer (3.1 mile) resolution, against roughly 5 to 10 kilometers for most major systems in routine use worldwide, the China Meteorological Administration said, per the South China Morning Post. The model runs on Sugon 8000, a 100,000 card system assembled entirely from domestically made AI accelerators that Sugon showed at the World Artificial Intelligence Conference in Shanghai in July. One 10 day run at that resolution finishes inside an hour, which the research team said meets international operational benchmarks. Researchers have also completed a global simulation test at 3 kilometer resolution. [South China Morning Post](https://www.scmp.com/news/china/science/article/3367747/china-uses-100000-home-grown-ai-chips-build-leading-weather-forecast-system) [Tencent News](https://news.qq.com/rain/a/20260915A07PEJ00) - **Xiaomi livestreams model training run with costs and chip failures on public view** — Xiaomi has opened a public dashboard streaming the reinforcement learning phase (training a model by scoring its own attempts) of its next model, MiMo-V2.6, showing money spent, steps completed, reward curves and which nodes have suffered chip failures, per QbitAI. The two variants had together cost more than $1.08 million in the 36 hours after the Pro model began training, averaging about $30,000 an hour. Luo Fuli, who leads the work, said the roughly six months since Xiaomi open sourced its previous model went into one question, how far reinforcement learning can be scaled. The three levers she named are training compute, the number of task environments a model works in, and the compute spent grading its attempts. Xiaomi's own offline scores on a software engineering benchmark put its two variants at 62.24 and 60.77, against 74.2 for DeepSeek's Flash model. [QbitAI](https://www.qbitai.com/2026/09/490950.html) # Policy Tracker - **Sen. Paul blocks unanimous consent for Sen. Kennedy's AI kill switch mandate** — Sen. Rand Paul (R-Ky.) objected on the Senate floor when Sen. John Kennedy (R-La.) sought unanimous consent to pass his AI bill, The Hill reported. The measure would require technology companies to build a kill switch for superintelligent AI. Paul said the requirement would amount to an open-ended government mandate, though Kennedy's bill would leave the shutdown control with the companies rather than the government. Kennedy said an AI bill would be impossible to pass before the midterms and likely through the lame-duck session, and described the kill switch as the only idea with a near-term path. [The Hill](https://thehill.com/homenews/senate/6094142-paul-kennedy-ai-kill-switch/) [Semafor](https://www.semafor.com/article/09/16/2026/congress-ai-safety-window-is-closing) - **AI bills would require enforcement machinery, New Yorker essay notes** — The enforcement machinery behind the AI bills now pending in Congress would reach into research practice itself, a New Yorker essay argues. Legislation from Sen. Bernie Sanders (I-Vt.) and Rep. Greg Casar (D-Texas) would make work on superintelligence punishable by as much as 20 years in prison and would stand up a new Cabinet-level federal agency for AI safety. The essay says such an agency would presumably have to surveil the work of computer scientists. It cites "AI 2040," a scenario from the nonprofit AI Futures Project where enforcement depends on monitoring hardware placed inside data centers, able to tell when a computer has begun training a model. The essay contrasts those measures with the pacing approach Anthropic chief executive Dario Amodei has proposed, which it says is neither stopping nor pausing. [The New Yorker](https://www.newyorker.com/culture/open-questions/the-long-doomsday-of-ai) [darioamodei.com](https://darioamodei.com/post/we-must-pace-the-frontier) - **White House shelves plan for industry-run AI regulator after executives called Trump** — White House aides paused work on creating a self-regulatory body for frontier AI after Trump soured on the idea, WIRED reported. Aides had started designing an oversight body attached to the administration soon after Google DeepMind co-founder Demis Hassabis floated the idea in a July 14 essay and in briefings to senior officials. The Treasury Department and the White House Office of Science and Technology Policy produced a preliminary draft under which the leading labs would police their own frontier work. Tech executives including former AI czar David Sacks and Meta chief executive Mark Zuckerberg called Trump one by one in August to object once they learned of the effort, leaving the proposal in limbo. No new AI oversight proposals have been seriously under consideration since. [WIRED](https://www.wired.com/story/washington-wont-be-regulating-ai-anytime-soon/) - **House passes Ratepayer Protection Act 417-3 to shift data center power costs** — The House passed the Ratepayer Protection Act 417-3, which directs state utility regulators to consider a standard requiring customers with peak demand of 100 megawatts or more, such as new data centers, to cover the full incremental cost of the generation, transmission and distribution upgrades needed to serve them, according to the Washington Post. Introduced by Reps. Gabe Evans (R-Colo.) and Kathy Castor (D-Fla.), the bill does not compel states to act and would not take full effect for two years. Evans said it also requires financial assurances that developers can finish projects instead of leaving half-built infrastructure for others to pay for, per Colorado Public Radio. Senate Majority Leader John Thune (R-S.D.) has signaled the Senate could take the bill up. [The Washington Post](https://www.washingtonpost.com/politics/2026/09/16/house-attempts-empower-states-lower-data-center-costs/) [The Hill](https://thehill.com/homenews/state-watch/6090232-house-passes-data-center-electricity-bill/) [House Energy and Commerce](https://energycommerce.house.gov/posts/ratepayer-protection-act-advances-from-house-committee-on-energy-and-commerce) - **NSF opens X-Labs competition for AI in physical systems under $1.5B program** — The National Science Foundation announced three more topics for NSF X-Labs, the $1.5 billion research program it began in May, completing its topic list for 2026. The newest funding opportunity invites proposals on artificial intelligence for physical systems, covering early stage work on advanced sensors, robotics and other machines that act in the physical world. The agency describes the labs as self-directed groups drawn from academia, engineering and startups that earn federal money by hitting milestones against a named scientific challenge. Two further topics are previewed for later this year, "From Sequence to Function" and "Computation at the Limit of Physics." [NSF](https://www.nsf.gov/news/nsf-announces-3-additional-topics-part-nsf-x-labs-initiative) - **Missouri Gov. Kehoe orders plate reader safeguards, urges ban on AI face matching** — Missouri Gov. Mike Kehoe signed an executive order on Sept. 16 directing the state Department of Public Safety to set statewide safeguards for law enforcement use of automated license plate readers, citing Flock Safety's cameras and data sharing network, StateScoop reported. The standards bind state agencies and state-funded law enforcement while the legislature works on a permanent statute, and other local agencies are encouraged to adopt them voluntarily. They include deleting most plate data within 30 days, barring vendors from selling data captured in Missouri, and prohibiting the integration of AI facial recognition with plate reader systems. Misuse would trigger disciplinary proceedings or criminal referral plus a mandatory report to the DPS director for possible license revocation. "The government should not be building a permanent record of where law-abiding Missourians go, who they visit, or how they live their daily lives," Kehoe said. [StateScoop](https://statescoop.com/citing-concerns-over-flock-cameras-missouri-governor-signs-executive-order-on-alprs/) [Missouri Governor](https://oembed-governor.mo.gov/press-releases/archive/governor-kehoe-signs-executive-order-establishing-safeguards-automated) # Capability & Research Watch - **Ex-Google researchers found London startup to build AI that improves itself** — Edward Hughes and Louis Kirsch left Google in December to found a London startup, Inherent, aimed at building an AI system capable of building a better AI system, the New York Times reported. Their prototype, Faraday, ingests the company's emails, instant messages, meeting transcripts and the researchers' own chats with Faraday, and Inherent uses that material to train a better version of it. Two Silicon Valley startups valued at $4 billion each are pursuing the same approach, and OpenAI and Anthropic are chasing it too. Anthropic said in a blog post this spring that its own pursuit of that capability could erode human control over AI systems. [The New York Times](https://www.nytimes.com/2026/09/16/science/ai-recursive-self-improvement.html) [Anthropic](https://www.anthropic.com/institute/recursive-self-improvement) - **Mathematicians confront AI after OpenAI's Navier-Stokes claim** — Mathematicians are reckoning with how quickly AI has entered their discipline, The Atlantic reports. OpenAI said last week that 10,000 of its agents had produced a solution to Navier-Stokes, a fluid flow problem and one of seven carrying the Millennium Prize, in under four days, building on nearly a century of human effort. In late July, UCLA professor Terence Tao told the International Congress of Mathematicians that AI is causing a "crisis in our mathematical values and practices." Earlier in that same conference week, an Anthropic researcher had used Claude to disprove a conjecture that had stood for 87 years. "We're seeing how swiftly an entire field can be ravaged by AI," Max Weinreich, a math professor at CUNY Baruch, told The Atlantic. [The Atlantic](https://www.theatlantic.com/technology/2026/09/math-crisis-openai-millennium-prize/688631/) # Industry & Market Watch - **Google, Nvidia and Emerald AI launch 20-member alliance for grid-flexible data centers** — Google, Nvidia and the startup Emerald AI launched the AI Energy Management Alliance, a coalition of 20 companies and organizations that will press regulators, grid operators and utilities to connect data centers faster when they commit to cutting power use as supplies tighten, per Axios. Launch partners include Anthropic, the utility National Grid and the power producer Constellation. The group is a relaunch of the Advanced Energy Management Alliance, founded in 2014 and largely inactive in recent years, which advocated for companies that trim electricity use when the grid is strained. The energy industry executive Frank Lacey will serve as executive director, per The Next Web. Flexibility would come from shifting or pausing computing jobs, discharging batteries or running onsite generation. [Axios](https://www.axios.com/2026/09/16/tech-giants-launch-flexible-power-coalition-data-centers) [The Next Web](https://thenextweb.com/news/ai-energy-management-alliance-google-nvidia-emerald-flexible-data-centres) [NVIDIA](https://blogs.nvidia.com/blog/ai-energy-management-alliance/) - **Basel Action Network puts AI data center waste at 23 million shipping containers by 2050** — Electronic waste from the AI buildout has been underestimated, according to a report from the nonprofit Basel Action Network, The Verge reported. The group projects the waste could fill 15 million to 23 million 40-foot shipping containers by 2050, enough to circle the globe about six times if lined up end to end. The estimate runs higher than earlier studies because the authors count the infrastructure that supports data center servers, not the servers alone. [The Verge](https://www.theverge.com/ai-artificial-intelligence/996470/ai-data-center-e-waste-ban) # Global & Geopolitics - **European Commission adopts KIDS Act proposal to put AI systems under child safety duties** — The European Commission adopted the EU KIDS Act, a proposal that would bar social media platforms from admitting children under 13 and set 15 as the bloc-wide floor for holding an account without a parent. The accompanying proposal document says the measure covers AI systems alongside social media services. It puts the burden on providers to demonstrate that what they offer suits a child's age rather than on regulators to prove harm. "Our KIDS Act is reversing the burden of proof; it is for platforms to show they are safe by design," Commission President Ursula von der Leyen said. The proposal must still pass the European Parliament and the Council of member state governments before it binds anyone. [European Commission](https://digital-strategy.ec.europa.eu/en/news/eu-kids-act-restrict-social-media-platforms-access-children-eu) - **Scottish Parliament votes to pause planning decisions on new AI data centers for up to a year** — Members of the Scottish Parliament rejected a Scottish Greens motion for a moratorium on new AI data centers and instead backed a Scottish Labour amendment saying no decisions on planning or consenting should be made until ministers publish national planning guidance, per the Guardian, following the campaign for a moratorium noted in AIPD's [Sept. 3 edition](https://aipolicydaily.org/archive/daily/2026-09-03/). The vote does not bind ministers, who have pledged guidance by the end of the year. Developers have put forward more than 20 large data centers in Scotland, among them a scheme at Auchtertool in Fife billed as the second largest in the world. The Scottish government had said before the vote that large-scale data centers would face mandatory environmental impact assessments. [The Guardian](https://www.theguardian.com/uk-news/2026/sep/16/datacentres-scotland-environmental-assessments-ai-boom) [The Scotsman](https://www.scotsman.com/news/politics/msps-back-ai-data-centre-planning-decision-pause-ahead-of-new-guidance-9068970) - **Spain's data regulator logs its first breach report blamed on an AI agent** — The Spanish Data Protection Agency (AEPD), the country's privacy regulator, said it received a breach notification attributing an attack to an AI agent running on a known large language model, BleepingComputer reported. The reporting organization said the agent searched for flaws in generic files, logged into the organization's systems, kept probing the application on its own, then altered personal data and accessed invoices. The agency has neither investigated nor verified the account, but said it shows AI related breaches are no longer theoretical. It said AI does not invent new threats so much as make attacks faster, larger and more adaptable. Procedures written for manual intrusions may be insufficient against agents that test access methods and adapt as they go, it said. [BleepingComputer](https://www.bleepingcomputer.com/news/security/spains-data-agency-gets-first-report-of-ai-powered-data-breach/) - **Applied Materials pledges $5B to India as Modi pitches alternative chip base** — Applied Materials committed $5 billion to India across the next 10 years, the U.S. chip equipment maker said at the SEMICON India conference in New Delhi, per Reuters. Prime Minister Narendra Modi used the conference, which drew more than 600 companies and representatives from 52 countries, to pitch India to manufacturers looking to spread chip sourcing beyond Taiwan and China. Bloomberg reported that Lam Research also pledged investment and put the scale of the government's push at $30 billion. India has committed more than $21 billion across two chip incentive programs, and commercial output from Tata Electronics' $10 billion plant in Gujarat is nearly two years late. Applied Materials' plan centers on a proposed research park near Bengaluru, per the Times of India. [Bloomberg](https://www.bloomberg.com/news/articles/2026-09-17/modi-courts-investors-with-india-s-30-billion-chipmaking-bet) [Reuters](https://www.reuters.com/world/asia-pacific/applied-materials-invest-5-billion-india-modis-flagship-chip-event-kicks-off-2026-09-17/) [The Times of India](https://timesofindia.indiatimes.com/business/india-business/applied-materials-plans-5-billion-india-investment-by-2035/amp_articleshow/134303689.cms)