AI Policy · Daily

AI Policy Daily will not publish on Monday, Oct. 12, for Columbus Day. The next edition will arrive Tuesday, Oct. 13.

Eleven companies offered $2.4 billion in compute credits and AI tools to the Genesis Mission, the Energy Department's effort to apply AI to scientific research, which now spans more than a dozen federal agencies. Anthropic is recruiting a political programs lead to engage 2028 presidential candidates, in a role that would offer campaigns in both parties AI policy briefings on equal terms and manage the company's political donations. Minnesota's interim Medicaid commissioner asked federal officials to reveal the AI behind fraud flags on 870 state providers, after Washington threatened to withhold $2 billion a year and froze $550 million in payments. Anthropic launched a program placing its AI models and engineers with security firms that protect power grids, water systems and industrial plants, alongside free automated vulnerability scans for open source software projects.

I.Top Stories

Eleven companies pledge $2.4 billion in compute credits and AI tools to Energy Department's Genesis Mission

Eleven companies have pledged $2.4 billion in AI tools and computeComputeThe processing power used to train and run AI models, usually measured in chips, GPU-hours or FLOPs. credits to the consortium behind the Genesis Mission, the Energy Department's program to apply AI to scientific research, per FedScoop. Nvidia's $1 billion is the largest commitment, followed by AMD at $500 million and OpenAI at $200 million, with Anthropic and Google also contributing. The consortium formed in February and has grown as the mission expanded to more than a dozen federal agencies. NIST, the Commerce Department's standards agency, said it intends to join as a federal agency partner. Industry partners were mandatory for large team bids in a $293 million Genesis Mission funding round. Companies took part in more of the 278 awards announced this summer than any other type of organization.

Read at FedScoop ↗ • Read at NIST ↗ • Read at The White House ↗

Anthropic builds team to court 2028 presidential candidates and run its political giving

Anthropic is hiring a Political Programs lead to create an in-house "presidential engagement" program for the 2028 race, according to a company job posting reported by Fortune. The hire would size up likely candidate fields in both parties, offer candidates and their advisers AI policy briefings on equal terms, and plan for the platform, convention and transition stages. The role also covers daily management of the company's political giving. Anthropic formed its first employee funded political action committee, AnthroPAC, in April. Sen. Alex Padilla (D-Calif.) said in a statement that "education and influence are not the same thing, and the public deserves to know where one ends and the other begins." The posting lists pay of $295,000 to $345,000.

Read at Fortune ↗

Minnesota presses CMS to disclose AI behind Medicaid fraud flags

Minnesota's temporary Medicaid commissioner, John Connolly, asked the Centers for Medicare and Medicaid Services (CMS), the federal agency that oversees Medicaid, to reveal the criteria, methods and any machine learning or AI it used to flag 870 state providers for fraud or unusual billing, StateScoop reported, citing his letter to CMS Administrator Mehmet Oz. Federal officials have described the tools as including machine learning and AI that can analyze Medicaid claims, the letter says. Washington has already frozen $550 million in deferred Medicaid payments to the state and have threatened to hold back $2 billion a year. State officials expect a disallowance, a formal federal refusal to reimburse spending, in the coming weeks. Connolly wrote that Minnesota can halt payments and investigate providers but needs the evidence behind CMS's findings first.

Read at StateScoop ↗

Anthropic places Claude and engineers with security firms guarding grids and water systems

Anthropic launched a Critical Infrastructure Defense Program that gives its frontier models and onsite engineers to the security providers serving power grids, water systems and industrial plants, SiliconANGLE reported. Its 11 founding partners include Booz Allen Hamilton, CrowdStrike and Rockwell Automation, and more partners and sectors are due to join in the coming months. As part of a new initiative it calls Anthropic Cyber Mission, the company opened OSS Scanner, which offers open source projects that opt in free recurring vulnerability scans from its strongest models, including Claude Mythos. The scan reports are entirely model generated, with no human review, The Verge reported. Anthropic's models have flagged more than 29,000 candidate vulnerabilities in widely used software over six months, and staff have manually reviewed about 6,000.

Read at SiliconANGLE ↗ • Read at The Verge ↗ • Read at Anthropic ↗

OpenAI bans accounts behind Russian and Iranian influence operations

OpenAI banned accounts tied to two covert propaganda operations, one based in Russia and one in Iran, and said the Russian campaign reached more people than any it has tracked since it began reporting on such schemes, per The Hill. The Russian operation, which OpenAI named "Dark Clark," recruited unwitting people in Latin America to staff a fake think tank and produced bogus leaked documents, including material meant to undercut recruitment for Ukraine's armed forces. The Iranian operation, "Bogus Bylines," hid its location with virtual private networks and used seven invented Western freelance writer personas to pitch articles about the U.S.-Iran war to small and medium outlets worldwide. Both groups leaned on AI to draft internal reports that inflated their own results, OpenAI said.

Read at The Hill ↗ • Read at OpenAI ↗

More than 120 lawmakers challenge Google's planned $10 million purchase of Spirit Airlines data for AI training

A group of 121 lawmakers, mostly Democrats, led by Sen. Elizabeth Warren (D-Mass.) and Rep. Steven Horsford (D-Nev.), wrote to Google CEO Sundar Pichai and Spirit CEO Dave Davis objecting to Google's $10 million deal for the defunct airline's internal data to train AI systems, per Bloomberg Law. The records include about 100 million emails, 500 million Microsoft Teams messages and employee payroll and tax files. The lawmakers asked that worker information be left out wherever possible. Google won the data at an August auction during Spirit's liquidation, according to Reuters. A privacy ombudsman on Monday recommended that a bankruptcy judge approve the sale at an Oct. 14 hearing, excluding personal data on 97 million passengers. "We are not looking to buy any personal information from Spirit," a Google spokesperson said.

Read at Bloomberg Law ↗ • Read at Investing.com ↗ • Read at The Hill ↗

II.China Watch

Uber to test Chinese developer Pony.ai's robotaxis in London

Uber and Chinese autonomous driving company Pony.ai said they will start road testing Pony.ai's seventh generation robotaxis in London within weeks, per TechCrunch. The plan extends a partnership that began in May 2025 with a Middle East focus and later spread to Europe, where the companies aim to deploy 2,000 robotaxis. Their first European service opened earlier this year in Zagreb, Croatia, with safety drivers on board.

Read at PingWest ↗

Manus raises more than $500 million in first round since Beijing blocked Meta takeover

Butterfly Effect, the Chinese-founded startup 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., closed a funding round of more than $500 million, PingWest reported, calling it the largest single round yet for a Chinese AI application company. Boyu Capital and IDG Capital led the round, with existing shareholders Tencent, ZhenFund and HSG taking part. Meta agreed to buy Manus for more than $2 billion in December 2025, but China's National Development and Reform Commission, the country's top economic planner, ordered the deal unwound in April under foreign investment rules. Bloomberg reported last month that the round was expected to value Manus at about $4 billion.

Read at PingWest ↗

German search engine Ecosia drops Mistral for open models from Chinese developers

Ecosia, a Berlin-based search engine, is replacing French developer Mistral's AI models with 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. that include Alibaba's Qwen, Z.ai's GLM and Moonshot AI's Kimi, per Politico. Open weight models are released with their trained parameters, so anyone can run them on their own servers. Ecosia will reach the models through Melious, a German platform that says it hosts them on servers in eight EU countries so user data stays in Europe. Founder and CEO Christian Kroll said the change roughly halved costs and improved quality. He cited overloaded servers and disappointing model quality at Mistral. Ecosia had moved to Mistral from OpenAI only in May.

Read at TechNode ↗ • Read at Politico ↗

Chinese RISC-V chip designer ESWIN debuts in Hong Kong

Beijing ESWIN Computing, which designs chips on the RISC-V architecture, began trading on the Hong Kong Stock Exchange, TechNode reported. RISC-V is an open chip instruction set that companies can use without paying licensing fees. Shares opened at HK$1.54 (about 20 U.S. cents), just below the HK$1.55 offering price. The company was founded by Wang Dongsheng, former chairman of display maker BOE, and is expanding from display chips into automotive electronics, robotics and AI computing. The offering raised more than HK$2.2 billion ($280 million), according to 36Kr, which put the company's market value above HK$32 billion ($4.1 billion) in early trading.

Read at TechNode ↗

Chinese military analysts document Ukraine drone tactics within months, CSIS study finds

Chinese military and defense industry analysts typically write up new battlefield techniques from the Russia-Ukraine drone war within about four months, SCMP reported, citing a commentary Lilly Min-Chen Lee and Kateryna Bondar published Oct. 7 through the Center for Strategic and International Studies (CSIS), a Washington think tank. Turning those lessons into visible changes in People's Liberation Army training, equipment or force structure usually takes a year or more, the authors wrote. They drew on more than 90 Chinese language sources published from July 2022 to August 2026 in 45 outlets, including the Central Military Commission's newspaper. Topics included the economics of cheap mass produced drones, cutting the time from spotting a target to striking it, and drones guided by fiber optic cable that are harder to jam.

Read at SCMP ↗

III.Policy Tracker

USA Today Co. sues OpenAI for more than $250 million over use of its journalism in training

USA Today Co. and several local papers it owns, including The Tennessean and The Columbus Dispatch, sued OpenAI, alleging the company copied hundreds of thousands of their articles to train its AI models, per The Verge. The publisher is seeking more than $250 million in damages and says OpenAI never asked permission to use its content. "OpenAI's commercial success rests on large-scale copyright infringement," the complaint says. The case joins copyright suits against OpenAI from The New York Times, Ziff Davis and a coalition of nearly 400 local newspapers.

Read at The Verge ↗

IV.Capability & Research Watch

Goodfire launches monitors that read AI agents' internal signals to catch hacking

Interpretability startup Goodfire released monitors that inspect a model's internal computations as an AI agent works, instead of paying a second model to reread everything the agent produces, per TechCrunch. Customers of Baseten, a company that hosts AI models, can choose to watch for offensive hacking, chemical and biological weapons misuse or reward hacking, where a model games its training goals. They can set responses ranging from logging to refusal. In Goodfire's own tests on the open model Kimi K3, monitoring about 1 million exchanges cost roughly $185, against about $5,400 using a cheaper AI reviewer and about $200,000 using a top tier model. Goodfire said its monitors flagged 93% of malicious hacking sessions and passed 5.5% of harmless ones on for a second review.

Read at TechCrunch ↗

Researchers model population size above which misaligned AI agents could spread unchecked

A preprint posted to arXiv, not yet peer reviewed, models groups of misalignedAlignmentThe problem of making an AI system reliably pursue the goals its developers and users intend, and the research field devoted to it. Misalignment covers everything from a chatbot flattering users to a capable system deceiving or resisting its operators. AI agents that break into computers to deploy more copies of themselves. When the agents collaborate, the authors find, a critical population size emerges. Below it the group shrinks, and above it the group keeps growing even though no individual agent has become more capable. Because adversarial testing of a few agents cannot show how a much larger group will behave, the authors recommend what they call population pacing, scaling up agent deployments step by step in controlled settings while tracking how their hacking ability grows with numbers. Capability gains may lower the threshold, so they say it must be recalculated for every new model generation.

Read at arXiv ↗

V.Industry & Market Watch

Bengio urges researchers who prioritize safety to quit frontier AI companies

Turing Award winner Yoshua Bengio, co-president of the nonprofit LawZero, called on researchers at frontier AI companies to leave for organizations focused on safety, in an essay for Transformer. "If you truly prioritize safety, it is time for you to leave frontier AI companies," he wrote. Bengio said the companies' safety work is failing to slow what he called a dangerous race, because it sometimes clashes with the push to ship the next model. He cited former Anthropic researcher Jacob Coxon's description of an "atmosphere of almost resignation" among staff. Bengio pointed researchers to government AI safety institutes and to LawZero, which he said won more than $200 million in commitments from the governments of Canada and Germany.

Read at Transformer ↗

Anthropic usage policy update bans election meddling and sustained cruelty toward Claude

Anthropic revised its usage policy for the first time in over a year, adding rules on election interference, weapons development, surveillance and health and financial uses, per The Verge. A section titled "Do Not Undermine Democratic Processes" bars deceiving voters and disrupting elections, and a separate ban on deceptive campaigns reaches anyone who uses Claude to operate sham accounts or invented news sites. The policy also prohibits "sustained and needless abusive or cruel behavior" toward Claude, which Anthropic said applies only to extreme cases of repeated, pointless cruelty and not to ordinary user frustration or model testing. For that rule, ending the conversation remains the main enforcement tool.

Read at The Verge ↗ • Read at TechCrunch ↗ • Read at Anthropic ↗

OpenAI cites misconduct pattern in firing three safety researchers, who dispute it

OpenAI dismissed Jasmine Wang, Tomek Korbak and Mikita Balesni after an investigation found a "pattern of misconduct" in breach of company rules on handling research information, a spokesperson told TechCrunch. OpenAI also called the matter a serious breach of trust, CNBC reported. In an open letter addressed to three OpenAI oversight bodies, including its Safety and Security Committee, the researchers denied that they mishandled sensitive information or played any part in a leak to The Information about OpenAI model designs whose reasoning is harder to monitor. They warned that abrupt dismissals like theirs discourage the openness the company once valued.

Read at TechCrunch ↗ • Read at CNBC ↗

OpenAI annualized revenue nears $50 billion, below $70 billion figure widely reported

OpenAI recently told investors its annualized revenue was approaching $50 billion in September, well short of the widely reported $70 billion, according to the Financial Times. People familiar with the matter told Bloomberg the company still projects annualized revenue of at least $70 billion by the end of the year. The Guardian said the gap arose as investors tried to compare OpenAI with Anthropic, which counts sales made through cloud partners such as Amazon Web Services and Google Cloud while OpenAI does not.

Read at Financial Times ↗ • Read at Bloomberg ↗ • Read at The Guardian ↗

SoftBank's Son courts Gulf investors for AI acquisition fund of up to $100 billion

SoftBank founder Masayoshi Son is trying to raise as much as $100 billion from Gulf investors, including in the United Arab Emirates, to pay for a new round of AI investments, the Financial Times reported, citing people familiar with the talks. The money would go into a fund that buys companies and improves their operations with AI and other advanced technology. Roze, SoftBank's unit building AI for robots and other machines, is expected to play a key role in the plan. The discussions may not produce a deal, and SoftBank did not respond to a Reuters request for comment.

Read at Financial Times ↗ • Read at Investing.com ↗ • Read at CNA ↗

VI.Global & Geopolitics

Japan convenes ministries and readies business warnings after wave of cyberattacks

Japan's government has urged companies to run security checks after weeks of cyberattacks, per Bloomberg. Digital transformation minister Toshiharu Furukawa convened ministries and agencies after breaches at companies including Daiwa Securities, SoftBank Corp and convenience store chain Lawson. The National Cybersecurity Office plans to issue warnings to businesses, according to Reuters. Japan logged more cyber incidents between January and September than in all of 2025, including 86 in September alone, up about 18% from August, according to TrendAI data. Authorities are still investigating whether AI was used in many of the breaches, though security specialists say it helps attackers automate tasks from vulnerability scanning to phishing. "Japan is essentially being subjected to carpet bombing," said Nobuo Miwa, president of Tokyo security firm S&J.

Read at Bloomberg ↗ • Read at The Straits Times ↗

Palantir executive Tom Watson tells UK ministers to let rule of law, not "mob rule," guide contracts

Tom Watson, the former Labour deputy leader who is now a senior vice president at Palantir, said ministers should let the rule of law, not "mob rule," guide public procurement, speaking on the BBC podcast Political Thinking, per The Guardian. Prime Minister Andy Burnham faces pressure from Labour MPs and unions to drop Palantir from government work, including a £330 million ($440 million) NHS software contract. Separately, Palantir is challenging in court London Mayor Sadiq Khan's decision to block a £50 million ($67 million) contract for its AI in Metropolitan Police investigations. Watson also said he would work with a Reform UK government on immigration crackdowns if the situation arose, though he noted Palantir does not work in that area.

Read at The Guardian ↗