AI Policy · Daily

President Trump pulled back an AI cybersecurity executive order hours before its expected signing, citing worries the draft could disadvantage U.S. firms against Chinese competitors. California Gov. Gavin Newsom ordered state agencies to examine AI linked labor disruption and recommend updates to the state's WARN Act within 180 days. The Coalition for Independent Technology Research filed suit against Secretary of State Marco Rubio over visa restrictions the group says chill foreign-born content moderation researchers.

I.AI Policy Today

Trump postpones AI cybersecurity executive order hours before Thursday signing, citing China competition concern

President Trump told reporters in the Oval Office Thursday he would delay his AI cybersecurity executive order because "I didn't like certain aspects of it," expressing concerns the language could harm U.S. AI competition with countries like China, CyberScoop reported. The cancellation reversed an expected Thursday signing, as reported by Bloomberg in AIPD's May 21st edition. The draft order would have required AI companies to share advanced models with the government between 14 and 90 days before launch. It would have given cybersecurity testers in critical infrastructure sectors such as finance and healthcare access to those models, authorized the NSA to conduct classified evaluations, and set up a Treasury information sharing agreement between AI companies and cybersecurity defenders. The Office of the National Cyber Director, CISA and NIST would also have defined which models the regime covered. TechCrunch reported the unofficial reason for the cancellation was that not enough tech CEOs could make it to Washington on short notice for the planned signing.

Read at CyberScoop ↗ • Read at TechCrunch ↗

California Gov. Newsom directs state agencies to study AI displacement and propose WARN Act updates within 180 days

California Gov. Gavin Newsom signed an executive order Thursday directing state agencies to study labor market shifts tied to AI adoption, including layoffs, hiring changes and skills gaps, StateScoop reported. The order calls for recommendations within 180 days on potential updates to California's WARN Act, which requires employers to provide advance notice of mass layoffs, to cover AI driven automation. Newsom said in a press release the state would not sit back and watch the future happen to it, and that the moment demands rethinking how Californians work, govern and prepare for the labor transition. The EO builds on Newsom's March order setting procurement and certification standards for AI vendors seeking state contracts, including transparency requirements and watermarking of AI generated content.

Read at StateScoop ↗

Coalition for Independent Technology Research sues Secretary Rubio over visa restrictions targeting online safety researchers

The Coalition for Independent Technology Research, an advocacy organization for technology researchers, sued Secretary of State Marco Rubio, former Homeland Security Secretary Kristi Noem and former Attorney General Pam Bondi over Rubio's visa restriction policy targeting foreign officials and other individuals deemed complicit in censoring Americans, MIT Technology Review reported. The complaint argues the policy violates the speech and due process rights of foreign-born technology researchers and workers whose work supports content moderation on technology platforms, and asks the court to halt the restrictions while the case proceeds. Knight First Amendment Institute senior staff attorney Carrie DeCell told reporters outside the Washington courthouse on May 13 that the policy is "expansive and incredibly vague, and the chilling effects are correspondingly enormous." The government filed a motion to dismiss the case at the end of last week.

Read at MIT Technology Review ↗

Transportation Department's Saini says AI agents must operate without human review to scale 500,000 FMCSA inspections

Ankur Saini, chief product and technology officer for several Department of Transportation units including the Federal Motor Carrier Safety Administration (FMCSA), said AI agents will need to operate outside the human-in-loop default to handle high-volume inspection work, FedScoop reported. "For these transformative use cases, I don't think there is an option but for the human to get out of the loop. Otherwise your human will always be your limiting factor," Saini said at ACT-IAC's Emerging Technology & Innovation Conference last week. FMCSA handles approximately 500,000 inspections per year across roughly 7 million drivers, with the agency concerned that fraudulent carriers are escaping current checks. AI agents could initiate documentation requests from motor carriers and take over low-risk inspection tasks.

Read at FedScoop ↗

II.China Watch

Zero Nvidia H200 chips sold in China since Trump's January approval

Not a single Nvidia H200 chip has been purchased in China since the Trump administration approved exports of the chip earlier this year, the New York Times reported, with Beijing steering domestic AI demand toward Huawei alternatives. Chinese customs agents were instructed in mid-January that H200s were not permitted to enter China, and Chinese government officials told domestic technology companies in meetings the same week not to purchase the chips unless necessary, Investing.com reported. Chinese technology firms had previously placed orders for more than 2 million H200 chips at around $27,000 each, far exceeding Nvidia's 700,000-chip inventory.

Read at NYT ↗ • Read at Investing.com ↗

DeepSeek raises 70 billion yuan as it confirms DeepSeek Code line

DeepSeek's latest funding round has reached 70 billion yuan (about $10 billion), with management telling potential investors the firm will prioritize breakthrough AI research over near-term commercialization, QbitAI reported. Senior researcher Deli Chen publicly posted recruitment notices referencing a "Harness team," "Code Harness" and "DeepSeek Code," confirming plans for a first party coding agent. Cui Tianyi, who joined DeepSeek in March, is expected to lead the Agent Harness team.

Read at QbitAI ↗

Alibaba's Qwen3.7-Max ranks fifth globally on Artificial Analysis index

Alibaba's flagship Qwen3.7-Max scored 56.6 on Artificial Analysis's latest global LLM leaderboard, placing fifth overall behind GPT, Claude and Gemini variants and finishing as the highest ranked Chinese model, QbitAI reported. The score outperformed Kimi-K2.6, DeepSeek-v4-Pro-Max and GLM-5.1, with most of the gain concentrated in programming, agent and reasoning benchmarks. Qwen3.7-Max will be made available via API on Alibaba Cloud's Bailian platform.

Read at QbitAI ↗

Huawei produces 122TB SSDs using proprietary Die-on-Board packaging

Huawei has produced 122-terabyte SSDs using its own Die-on-Board (DoB) packaging, a wafer level approach that mounts NAND dies directly onto a printed circuit board for a 33% capacity density gain over conventional packaging, Pandaily reported. The technique sidesteps Huawei's lack of access to advanced 3D NAND from mainstream non-Chinese suppliers, with the company leaning on domestic NAND from YMTC and other Chinese makers. Huawei outlined a roadmap including 61.44TB and 122.88TB drives already in production, with a 245TB version planned.

Read at Pandaily ↗

III.Federal Policy Tracker

FTC requires Cox Media Group and two marketing firms to pay $930,000 over false "Active Listening" AI claims

The Federal Trade Commission will require Cox Media Group and two smaller marketing firms to pay a total of $930,000 to settle allegations they deceived customers by falsely claiming an AI powered service could target localized advertising using conversations captured from consumers' smart devices, the agency said in a press release. The three complaints named CMG Media Corporation (doing business as Cox Media Group), New Hampshire-based MindSift LLC and Wisconsin-based 1010 Digital Works LLC. The FTC alleged the firms claimed an "Active Listening" algorithm could listen in on conversations overheard by smart devices and target ads within a desired geographic region. The service did not use voice data and instead resold marked-up email lists obtained from other data brokers. Christopher Mufarrige, director of the FTC's Bureau of Consumer Protection, said the companies misled potential customers by falsely claiming consumers had opted in.

Read at FTC ↗

GSA strikes OneGov deal with Snowflake, offering federal agencies 20% off compute services

The General Services Administration announced a OneGov contracting deal with Snowflake making the company's data and AI cloud products available to all federal agencies, FedScoop reported. Federal users will receive 20% off Snowflake compute services, scaling to 50% with usage, plus nearly 27% off storage. Snowflake gained FedRAMP authorization for AWS GovCloud in 2023 and for Microsoft Azure Government last year, and joins OpenAI, Anthropic and Perplexity tools already available through OneGov. The Snowflake agreement runs through Sept. 30, 2027.

Read at FedScoop ↗

Pentagon CDAO launches 180-day GenAI.mil Task Force to embed AI experts in operational units

The Pentagon's Office of the Chief Digital and Artificial Intelligence Officer established the GenAI.mil Task Force (GenTF) as a 180-day pilot program to embed AI experts within operational units to support generative AI integration into mission workflows, ExecutiveGov reported. CDAO appointed Air Force Capt. Anthony McHugh and Army Capt. Ryan Hetrick as task force co-leads; Hetrick previously served as Future Operations and AI Lead within the Cyber National Mission Force. GenAI.mil, launched in December, surpassed 1 million unique users within its first two months and offers Google's Gemini for Government, xAI for Government and OpenAI's ChatGPT.

Read at ExecutiveGov ↗

Federal prosecutors charge two men under Take It Down Act for creating AI generated celebrity nudes

U.S. federal prosecutors charged Cornelius Shannon, 51, and Arturo Hernandez, 20, with using AI to create nude videos and photos of female celebrities under the newly enacted Take It Down Act, AP reported. The two men, who do not appear to be connected, were arrested Tuesday and charged Thursday, with their content drawing millions of views online per the criminal complaints. The Take It Down Act, signed by Trump last year with bipartisan support and Melania Trump's public backing, adds stricter penalties for publishing AI generated deepfakes and revenge porn; the men face up to two years in prison. U.S. Attorney Joseph Nocella in Brooklyn said in a statement that the men used digital tools to create images that degraded and violated dozens of women.

Read at AP ↗

Commerce signs $2 billion in CHIPS Act quantum awards plus noncontrolling equity stakes in nine companies

The Commerce Department signed letters of intent to distribute roughly $2 billion to nine quantum computing companies under CHIPS Act microelectronics provisions, with a noncontrolling equity stake conditioning each award, FedScoop reported. IBM will receive $1 billion to back a new standalone subsidiary, Anderon, slated to operate as a 300-millimeter quantum wafer foundry; GlobalFoundries will receive $375 million, with Commerce taking approximately 1% equity, to launch a Quantum Technology Solutions business. Atom Computing, Diraq, D-Wave, Infleqtion, PsiQuantum, Quantinuum and Rigetti each received between $38 million and $100 million. Secretary of Commerce Howard Lutnick said the investments would build on the U.S. domestic industry and create thousands of high-paying American jobs.

Read at FedScoop ↗

IV.Capability & Research Watch

UK AISI report warns AI oversight properties face "severe degradation pathways" without intervention

A report by the UK AI Security Institute's Model Transparency and Situational Awareness teams warns that many properties relied on for current AI oversight face potentially severe degradation pathways absent effective intervention, based on interviews with AI safety researchers and AISI's own analysis, per a LessWrong summary. Today's oversight regime relies on chain-of-thought monitoring, white-box analysis and prefill attacks to detect misaligned model behavior and hidden propensities. The report recommends specific measurements for oversight-relevant properties, work to preserve oversight and investment in emerging oversight techniques as fallbacks.

Read at LessWrong ↗

Kapoor-Narayanan paper proposes open-world AI evaluations, tests agent on iOS app submission as first case

An arXiv paper from Sayash Kapoor, Helen Toner, Arvind Narayanan and 15 co-authors proposes "open-world evaluations" as a complement to benchmark-based testing for frontier AI capabilities. The authors argue benchmarks can both overstate and understate deployed capability by privileging tasks that are precisely specified, automatically graded and easy to optimize for. As a first instance, the paper tasked an AI agent with developing and publishing a simple iOS application to the Apple App Store, with the agent completing the task with a single avoidable manual intervention. The paper introduces CRUX (Collaborative Research for Updating AI eXpectations), a project for conducting recurring open-world evaluations.

Read at arXiv ↗

V.Industry & Market Watch

Spotify-Universal Music agreement opens UMG catalog to AI covers and remixes for premium subscribers

Spotify and Universal Music Group agreed to a licensing deal allowing premium Spotify subscribers to generate AI covers and remixes of UMG catalog songs, The Guardian reported. The deal will be structured as a paid add-on to Spotify Premium with revenue share for participating artists, in addition to existing royalty payments. Spotify co-CEO Alex Norström said the design centers on consent, credit and compensation for participating artists and songwriters. UMG home artists include Taylor Swift, Ariana Grande and Billie Eilish; Spotify shares closed up 16% Thursday following the announcement and accompanying revenue outlook.

Read at The Guardian ↗

Cloudflare cuts 20% of workforce despite record revenue, with middle management bearing brunt

Cloudflare CEO Matthew Prince said the company cut 20% of its workforce earlier this month despite posting record revenue, in a Wall Street Journal op-ed published Wednesday and reported by Fortune. Prince framed the cuts around "measurers," whom he defined as middle management, finance, legal, internal auditing and revenue recognition staff. The company will prioritize builders and sellers in future hiring, Prince wrote. Cloudflare operated with more than 5,000 employees as of 2025 across 13 global offices in a hybrid model.

Read at Fortune ↗

OpenAI's Lehane wants AI societal-impact debate softened and state laws that don't slow growth

OpenAI global affairs chief Chris Lehane told Wired this week that he wants to "tone down the debate" over AI's societal impacts and to persuade state lawmakers to adopt regulations that won't hamper the company's growth. Lehane described public narratives about AI as artificially binary, characterizing one camp as expecting a future where nobody has to work and the other as expecting a small group of elites controlling all-powerful AI. Lehane joined OpenAI in 2024 after working at Airbnb and as a founding force behind Fairshake, the crypto industry super PAC that worked to legitimize digital currencies in Washington. He previously served in Bill Clinton's White House on crisis communications.

Read at Wired ↗

VI.Global & Geopolitics

Spain's economy minister says EU-Anthropic Mythos testing talks have stalled

Spain's economy minister Carlos Cuerpo told Bloomberg the European Union has made little progress in talks with Anthropic over getting EU banks and companies tested for digital vulnerabilities surfaced by the Mythos AI model. Cuerpo earlier called for early EU-level access to Mythos via the European Commission to avoid becoming a second tier region on cyber defense, per Investing.com. Anthropic has restricted Mythos to a select group of tech companies and banks including JPMorgan and Apple, with the company also briefing the UK chaired Financial Stability Board, The Guardian reported. UK AISI's most recent appraisal of Mythos found it solved a previously unsolved cybersecurity benchmark known as "cooling tower" in 3 of 10 attempts.

Read at Bloomberg ↗ • Read at Investing.com ↗ • Read at The Guardian ↗

London Mayor Sadiq Khan blocks £50M Met Police AI deal with Palantir, citing procurement rule breaches

London Mayor Sadiq Khan blocked a £50 million ($66 million) deal between the Metropolitan Police and Palantir Technologies, citing serious breaches of procurement rules, The Guardian reported. The contract would have used Palantir AI to automate intelligence analysis in criminal investigations and would have been Palantir's largest UK policing deal to date, after £330 million ($430 million) NHS England and £240 million ($310 million) Ministry of Defence deals. The Mayor's Office for Policing and Crime told the Met that police had seriously considered only one supplier, Palantir. Scotland Yard said the decision was disappointing and said it could affect policing capacity.

Read at The Guardian ↗

EU lawmakers reach provisional agreement banning AI nudifier apps and AI generated CSAM tools

EU lawmakers reached a provisional agreement banning AI nudifier applications and systems used to generate child sexual abuse material, with companies given until Dec. 2 to comply, the National Catholic Register reported. Irish MEP Michael McNamara, a lead lawmaker on the AI Act, told EWTN News negotiators pushed for the prohibitions to sit in Article 5 of the AI Act among the absolute bans. The legislation also postpones some "high-risk" AI system obligations to 2027 and 2028, with lawmakers citing delays in implementation standards. The provisional agreement lands days ahead of Pope Leo XIV's first encyclical, "Magnifica Humanitas," on the protection of human dignity in the age of AI, scheduled for release May 25.

Read at National Catholic Register ↗