AI Policy · Daily

Sens. Banks and Cotton asked the intelligence community to benchmark U.S. against Chinese AI capability and brief Congress on the results. India began stress testing Aadhaar, Finacle and CERT-In systems against Anthropic's Mythos model, with Infosys and Tata Consultancy Services running the work in a secure environment. Taiwan prosecutors charged three people with routing Nvidia equipped Super Micro servers to China through Japan using forged export paperwork. President Trump endorsed exclusive Commodity Futures Trading Commission (CFTC) jurisdiction over prediction markets as Minnesota's new criminal penalties drew a federal lawsuit from the agency.

I.AI Policy Today

Sens. Banks and Cotton press IC to refocus assessment on China AI capabilities

Sens. Jim Banks (R-Ind.) and Tom Cotton (R-Ark.) sent a letter to Director of National Intelligence Tulsi Gabbard, CIA Director John Ratcliffe and the National Cyber Director asking the intelligence community to focus assessment efforts on China's AI capabilities, The Hill reported. The letter was shared first with The Hill. Banks and Cotton ask the IC to compare U.S. and PRC AI capability directly and to brief Congress on the findings.

Read at The Hill ↗

India tests Aadhaar, Finacle and CERT-In digital infrastructure for Mythos vulnerabilities

India is running stress tests of its most sensitive public-facing financial and government application software against Anthropic's Mythos AI model, Bloomberg reported, citing Indian officials familiar with the matter. Infosys and Tata Consultancy Services are conducting the tests in a secure environment, with Infosys focused on patches for its Finacle banking software used by financial firms globally. CERT-In is separately testing the Aadhaar national ID program and government login systems, using Anthropic's Claude Opus 4.7 to patch flaws because the companies do not have direct Mythos access. New Delhi has identified a domestic facility for in-country Mythos testing and the Ministry of External Affairs is leading those talks with U.S. officials, according to Indian officials.

Read at Bloomberg ↗ • Read at Business Standard ↗

Taiwan prosecutors charge three with smuggling Nvidia AI servers to China via Japan

At least one shipment of Super Micro servers containing advanced Nvidia chips was successfully smuggled to China via Japan before Taiwanese authorities could intervene, Bloomberg reported, citing people familiar with the matter. Three individuals were detained for using forged export documents to route the servers through Japan and on to Hong Kong. The case follows March charges against a Super Micro co-founder Yih-Shyan Li and two associates accused of conspiring to smuggle Nvidia-powered servers to China.

Read at Bloomberg ↗

Trump backs exclusive CFTC authority over prediction markets as state cases proliferate

President Trump said in a Truth Social post Tuesday that it was "critically important" for the Commodity Futures Trading Commission to maintain "exclusive authority" over prediction markets, The Hill reported. Trump named Chris Christie, New York Attorney General Letitia James, Minnesota Gov. Tim Walz and Illinois Gov. JB Pritzker among the present and former state officials trying to regulate the platforms. Minnesota last week became the first state to enact criminal penalties for operating prediction markets, and the CFTC filed a federal lawsuit to overturn the law the day after Gov. Walz signed it. Weekly trading volume on market leader Kalshi has risen from about $100 million last year to more than $3 billion, according to a CNBC estimate cited by the Guardian.

Read at The Hill ↗ • Read at The Guardian ↗

Vermont legislature sends Gov. Scott a bill drawing a line between administrative and clinical AI in mental health

Vermont lawmakers gave final approval Tuesday to H.816, a bill that limits AI in mental health care by distinguishing administrative uses from clinical decision-making, Seven Days reported. The measure cleared a conference committee last week and passed both chambers on Tuesday before heading to Gov. Phil Scott for signature. The bill statutorily defines which AI uses are permitted in clinical mental health treatment and which require a licensed provider. Scott's signature would put Vermont among the first states to enact a statutory administrative/clinical AI line in mental health care.

Read at Seven Days VT ↗

Vance calls Pope Leo's AI warnings "profound," opening visible split with rest of Trump senior team

Vice President JD Vance on Tuesday called Pope Leo XIV's AI warnings "profound," producing a visible split inside President Trump's senior team. The rest of the White House continues to resist new AI guardrails, CNBC reported. Interior Secretary Doug Burgum took the opposite line, defending the administration's deregulatory posture. The encyclical, "Magnifica Humanitas," calls AI use in lethal autonomous decisions "not permissible." Anthropic co-founder Christopher Olah appeared on stage with the Pope at the Monday presentation in Vatican City.

Read at CNBC ↗ • Read at The Hill ↗

Trump seats former AG Pam Bondi on White House AI advisory committee

President Trump appointed former U.S. Attorney General Pam Bondi to a White House advisory committee focused on AI, per Axios. Bondi was ousted as attorney general last month. The Axios report says the committee covers AI development, AI regulation, cybersecurity and the impact of emerging technologies on the U.S. economy and national security.

Read at Axios ↗

II.China Watch

Qualcomm strikes ASIC supply deal with ByteDance for AI data centers

Qualcomm has agreed to supply ByteDance with millions of custom AI chips, known as ASICs, to power AI agent workloads in the TikTok parent's data centers, Pandaily reported. The arrangement converts a ByteDance in-house chip design into a production ready semiconductor and gives Qualcomm one of its first large-scale customers for AI accelerators. Bloomberg first disclosed the deal on May 26, sending Qualcomm shares up more than 5%. U.S. export controls so far have focused on Nvidia and AMD GPU shipments, not custom ASIC flows.

Read at Pandaily ↗

Alipay's AI agent payments cross 300 million as Ant Group launches "AI Wallet"

Ant Group's Alipay said its AI Pay service has processed more than 300 million transactions to date, Pandaily reported. At the Alipay AI Payment Ecosystem Conference in Hangzhou on May 26, the company unveiled an AI Wallet that lets users authorize and audit purchases executed by AI agents, plus a Token Pay platform for AI model subscriptions and usage billing. Chinese model developers MiniMax and Stepfun signed on as early Token Pay partners. The rollout puts agentic commerce inside one of the world's largest payments rails.

Read at Pandaily ↗

AI demand drives April industrial profits up 24.7%, fastest growth since 2023

Profits at China's industrial enterprises above designated size grew 24.7% year over year in April, the fastest pace since December 2023, the National Bureau of Statistics said Wednesday, Caixin reported. Electronics industry profits surged 107.7% in the first four months, contributing 43.8% of all industrial profit growth, while equipment manufacturing and consumer goods profits broadly declined outside electronics. The article attributed the acceleration to surging AI related demand and higher oil prices feeding chemicals and nonferrous metals.

Read at Caixin ↗

ChinAI calls Anthropic's "2028" paper dogmatic, disputes diffusion assumptions

Jeffrey Ding's ChinAI newsletter on Tuesday published a critique of Anthropic's "2028: Two Scenarios for Global AI Leadership" paper, calling it "dogmatic" and identifying seven unfounded assumptions. Ding, a George Washington University political scientist, disputes the paper's claim that military AI capabilities will diffuse quickly and can deploy in weeks. He argues that general purpose technologies historically diffuse over decades and that China's cloud adoption sits at roughly half the U.S. rate. He also challenges the paper's binary framing of U.S. dominance versus authoritarian AI rule, noting that the AI stack is already interconnected through arrangements like Perplexity offering subscribers access to Moonshot's Kimi K2.5 model. Anthropic's "2028" paper is among the documents that endorse the AI Overwatch Act compute flow controls.

Read at ChinAI Newsletter ↗

III.Federal Policy Tracker

Commerce grants Volvo a carve-out from the Connected Vehicle Rule

The Trump administration gave Volvo Cars, owned by Geely, specific authorization to keep importing and selling connected vehicles in the U.S. despite rules targeting Chinese developed vehicle software and hardware, TechCrunch reported. Volvo said the approval followed discussions with Commerce on the company's governance, technology and data security. The Biden era rule blocks Chinese developed software starting with 2027 model-year vehicles and Chinese connected hardware starting with 2030 model-year vehicles. Volvo's sister company Polestar, also majority owned by Geely, said Tuesday it continues to work with U.S. authorities to meet the requirements.

Read at TechCrunch ↗ • Read at Investing.com ↗

CISA orders federal agencies to patch actively exploited Drupal SQL flaw by Wednesday evening

The Cybersecurity and Infrastructure Security Agency gave federal agencies until Wednesday evening to secure Drupal CMS instances against an actively exploited SQL injection vulnerability added to its Known Exploited Vulnerabilities catalog, BleepingComputer reported. Drupal is widely used by federal agencies, research universities and media organizations, broadening the patch population the directive applies to. CISA continues to issue same-week KEV directives even as Axios reported staff and funding cuts have diminished the agency's capacity.

Read at BleepingComputer ↗ • Read at Axios ↗

IV.Capability & Research Watch

Stanford-led study finds AI hiring tool produced adverse impact in 10.6% of jobs analyzed

A Stanford, Chapman and Northeastern team analyzed 4 million applications from 3 million applicants across 156 employers using Pymetrics, finding adverse impact against Black applicants in 10.62% of the 1,746 individual positions, the Financial Times reported. About 25.87% of all applications from Black applicants, nearly 40,000 submissions, went to positions where the algorithm produced outcomes that triggered federal four-fifths rule scrutiny. Researchers also documented "systemic rejection": 4% of applicants who applied to 10 Pymetrics screened positions were rejected from every one because the same score is reused across employers for up to 330 days. The paper, "Algorithmic Monocultures in Hiring," will be presented at the ACM Conference on Fairness, Accountability, and Transparency in Montreal next month.

Read at Financial Times ↗ • Read at Fortune ↗

"BadHost" flaw in Starlette exposes servers hosting millions of AI agents to credential theft

Security researchers disclosed a critical vulnerability called "BadHost" in Starlette, a Python web framework that serves as the runtime for many AI agent deployments and receives roughly 325 million downloads per week, Ars Technica reported. The flaw allows attackers to breach servers running AI agents and steal credentials to third-party accounts the agents use. Thousands of other open source projects depend on the vulnerable package.

Read at Ars Technica ↗

V.Industry & Market Watch

Altman softens AI "jobs apocalypse" warnings, says he is "delighted to be wrong"

OpenAI CEO Sam Altman walked back his white-collar wipeout warnings at a Commonwealth Bank of Australia conference, saying he doesn't expect "the kind of jobs apocalypse that some of the companies in our space advocate," Time reported. Altman said he is "delighted to be wrong" on earlier predictions that entry-level roles would be gone by now. AIPN's Peter Wildeford told Time it's "hard to say whether they've actually changed their forecasts for AI's economic impact, or whether they're just trying to change the narrative." A May Yale Budget Lab study found no meaningful unemployment shift through March in high-exposure jobs.

Read at Time ↗

VI.Global & Geopolitics

ECB warns AI boom is feeding private credit risk, with insurers and pensions most exposed

The European Central Bank's financial stability review extract released Tuesday added AI investment exposures to its private credit risk list, warning that insurers and pension funds would be hit hardest by a severe private credit shock, Insurance Journal reported. Euro area insurers have about 211 billion euros (about $230 billion) of exposure to private credit, or 2.3% of total assets, while pension funds hold 52 billion euros (about $57 billion), or 1.4%. Banks' losses would be contained at no more than 1.3% of total equity. Investor concerns about AI overspending and AI's disruptive effect on the software sector were among the drivers of recent private credit market turbulence.

Read at Insurance Journal ↗ • Read at Financial Times ↗

Germany and Spain push back on European Commission plan to ban Huawei from telecom networks

Germany and Spain are leading member state opposition to a European Commission proposal to ban Chinese telecom suppliers from EU networks under a new Cybersecurity Act, Bloomberg reported. The European Commission earlier this month told member states to phase Huawei and ZTE hardware out of connectivity infrastructure, and industry groups including the GSMA and Connect Europe have called the proposed remedies "unnecessary and disproportionate." Under a German compromise, Chinese vendors are being removed from core networks and from critical software in the radio access network while antennas and other hardware remain in use. China last week threatened retaliation if the rules take effect.

Read at Bloomberg ↗

GCHQ director warns of narrowing window to stay ahead of Chinese AI capabilities

GCHQ Director Anne Keast-Butler will say in a Bletchley Park lecture Wednesday that Russia is "relentlessly targeting critical infrastructure, democratic processes, supply chains and public trust" in Britain, and that AI advances mean a narrowing window for the UK and its allies to stay ahead of China, AP reported. Britain faces four major cybersecurity incidents a week, with China, Russia and Iran behind most of the serious attacks, according to figures issued last month by NCSC Chief Executive Richard Horne. Keast-Butler is calling for an effort to make cybersecurity 10 times more urgent. Keast-Butler also warned that "the risk of miscalculation" with state adversaries is the highest she has ever seen, The Guardian reported.

Read at AP ↗ • Read at The Guardian ↗

South Africa's DCDT sets November Cabinet target after AI hallucinations sank prior policy draft

South Africa's Department of Communications and Digital Technologies has set a fiscal 2026/2027 target for a revised national AI policy, aiming to present an updated draft to Cabinet by November and open it for public comment by January 2027, ITWeb reported. Communications and Digital Technologies Minister Solly Malatsi pulled the original policy after AI hallucinations produced fictitious academic citations and authors in the draft. An expert panel has been convened to lead the rewrite.

Read at ITWeb ↗ • Read at Citizen ZA ↗