AI Policy · Daily

We will not publish AIPD tomorrow due to the federal holiday. Please enjoy the long weekend and see you Monday!

The White House made blocking all jailbreaks its condition for restoring Anthropic's Fable 5 and Mythos 5 exports, a bar researchers say no frontier system can clear, even as Anthropic employees say the company is being singled out. Dario Amodei, Demis Hassabis and Sam Altman backed a U.S.-led AI coalition at the G7 lunch in Evian-les-Bains, with Altman floating a global standards forum, as Trump told reporters the Anthropic talks are "going fine." Democratic operatives launched the Guardrails Alliance, a $5 million super PAC built to counter Leading the Future, the pro-AI group whose budget tops $100 million, ahead of the midterms. A separate White House directive forced Anthropic to cut SK Telecom from Claude Mythos over alleged Chinese ties, days before the broader export takedown landed.

I.Top Stories

White House makes blocking all jailbreaks the bar for restoring Anthropic's Fable 5 and Mythos 5

Trump administration officials told Wired that Anthropic must address the alleged vulnerabilities in Fable 5, the guardrailed version of Mythos 5, before the export controls that took the models offline are lifted. Anthropic has called the concerns overblown and the jailbreak effects minimal, a position it reiterated to the Commerce Department and National Cyber Director Sean Cairncross in a Monday technical meeting, while independent cybersecurity experts say preventing all jailbreaks on frontier models is not possible. Separately, Anthropic employees told the New York Times they feared being singled out, noting that rival firms with comparable models have drawn no similar scrutiny, after the order suspended access to both models for all foreign nationals.

Read at Wired ↗ • Read at NYT ↗

Guardrails Alliance launches to counter pro-AI super PAC money in the midterms

Democratic operatives Shaunna Thomas and Leah Hunt-Hendrix unveiled a super PAC Thursday, the Guardrails Alliance, backed by tech workers, labor unions and left-leaning groups, the New York Times reported. The group has raised $5 million across the PAC and a linked nonprofit, with a goal of $15 million this cycle, and casts itself as a populist, small-dollar counterweight to Leading the Future, the super PAC favoring lighter AI regulation, whose political budget tops $100 million. Guardrails said its backers include the American Federation of Teachers, the Working Families Party, former Indeed CEO Chris Hyams and former OpenAI researcher David Farhi.

Read at NYT ↗

Trump administration shapes AI through a 'shadow' policy with no formal rules

Despite entering office vowing to get government out of the AI industry's way, the White House is steering the sector through case-by-case interventions rather than regulation, Axios reported. Export controls, voluntary testing frameworks and procurement guidelines have become the de facto building blocks of its approach, operating outside the formal rulemaking process and leaving companies little published guidance. With Congress frozen on AI ahead of the midterms despite a bipartisan House safety bill, the administration is taking the lead through executive action.

Read at Axios ↗

White House ordered Anthropic to revoke SK Telecom's Mythos access days before broader export action

A separate White House directive in early June told Anthropic to cut off South Korean telecom giant SK Telecom from Claude Mythos over alleged ties to Chinese entities, Wired reported. SK Telecom had received Mythos access through Anthropic's Project Glasswing rollout earlier in the month and immediately lost it after Anthropic complied with the order. Washington did not threaten broader export controls on the model at that time, with the public takedown order following days later.

Read at Wired ↗

Amodei, Hassabis and Altman back U.S.-anchored AI coalition at G7 lunch as Trump says Anthropic talks 'going fine'

Anthropic CEO Dario Amodei and Google DeepMind CEO Demis Hassabis endorsed a U.S.-led international AI coalition during the G7 working lunch in Evian-les-Bains, with OpenAI CEO Sam Altman supporting the same framing, per CNBC. Altman, who sat between Trump and Egyptian President Abdel Fattah el-Sisi, proposed an international forum modeled on the Financial Stability Board to set global AI standards. Canadian Prime Minister Mark Carney, who chaired the FSB from 2011 to 2018, likened the proposed forum to the FSB, per Semafor. Treasury Secretary Scott Bessent, Commerce Secretary Howard Lutnick and Secretary of State Marco Rubio also attended. Speaking to reporters before the lunch, Trump said negotiations with Anthropic are "going fine," per WSJ.

Read at CNBC ↗ • Read at Semafor ↗ • Read at WSJ ↗

II.China Watch

Zhipu advances its Shanghai IPO bid as a new flagship model lifts its Hong Kong cap past HK$930 billion

Zhipu AI cleared coaching acceptance, the sponsor sign-off step in China's pre-IPO process, on Wednesday. Guotai Haitong Securities is sole sponsor after CICC stepped back, Caixin reported. The Beijing-based lab also released a new open source modelOpen-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., GLM-5.2, that day, lifting Hong Kong shares 12.6%, with a further 26.1% gain Thursday to a close of HK$2,094 per share (about $268) that pushed market cap above HK$930 billion (about $119 billion). GLM-5.2 placed second on Code Arena, a UC Berkeley blind test leaderboard, behind only Anthropic's Claude Fable 5, the same model Washington blocked for foreign nationals on Friday. (Caixin)

Beijing's state communications group tells Lhasa forum that winning the Tibet narrative means adapting to Western AI feeds, not fighting them

At the Second Xizang International Communication Conference in Lhasa on Tuesday, Zachary Lundquist of the state-run China International Communications Group said the biggest obstacle to projecting Beijing's image of Tibet was no longer political hostility but the "invisible, self-reinforcing recommendation engines of Silicon Valley," the South China Morning Post reported. Lundquist, who uses the Chinese name Huang Hao, said Western searches for "Tibet" tend to surface a "pre-existing, highly politicised narrative," even though Beijing has used the romanized name "Xizang" since 2023 to assert its sovereign framing. He addressed more than 300 media professionals, officials and academics at the state-organized event.

Read at South China Morning Post ↗

Chinese state telecom giants' AI token plans flop with developers as alternatives undercut them by sixfold

China Mobile, China Unicom and China Telecom have rolled out AI token packages starting at 9.9 yuan ($1.40) for 10 million tokens, but testing by National Business Daily found a single "hello" prompt burning through about 50,000 tokens, per Pandaily. One developer estimated that daily traffic of 61.7 million tokens would cost around 1,000 yuan ($140) a month under operator pricing, against 149 yuan ($21) for his current Zhipu GLM subscription. Reporters who visited operator service halls in Beijing, Shanghai, Chengdu and Chongqing found that frontline staff did not know the packages existed, with one China Unicom employee calling them "just a concept." Omdia senior analyst Maosen Xia said operators "lack pricing power in the Token value chain" and will remain computing powerComputeThe processing power used to train and run AI models, usually measured in chips, GPU-hours or FLOPs. providers rather than AI service platforms.

Read at Pandaily ↗

Haier wraps factory operations into an enterprise 'Agent OS' as Beijing pushes manufacturing AI rollouts

Appliance maker Haier Smart Home has built an internal agent platform, Zhixiaoneng, that serves as the single AI entry point for employees handling orders, inventory, R&D, sourcing, marketing and customer service, 36Kr reported. Tencent Cloud is the technical partner. CTO Cui Xiuyuan said the system serves as the company's enterprise Agent OS, with AI now deployed across 9 business areas, 5 platforms and more than 70 scenarios spanning research, production, supply and service. Employees have built more than 5,100 lightweight applications and 262 self-made agents through the platform.

Read at 36Kr ↗

III.Federal Policy Tracker

NRC Chair Nieh details reactor licensing overhaul as AI grid load pressures regulator

Nuclear Regulatory Commission Chair Ho K. Nieh told Bloomberg the agency has taken major steps over the past 12 months to reform its reactor licensing framework in response to grid load pressure from AI data centers. Nieh said the NRC issued a new technology-neutral licensing framework in March to speed reviews for advanced reactor designs, including those using novel cooling methods and fuels. He said the NRC is not lowering safety standards.

Read at Bloomberg ↗

Pentagon AI chief's sworn declaration on Grok's Iran role anchors DOJ defense of xAI's Mississippi power plant

Cameron Stanley, the Pentagon's chief digital and AI officer, filed a sworn declaration Monday saying Elon Musk's Grok helped U.S. forces deploy more than 2,000 munitions on 2,000 distinct targets in 96 hours during the Iran war, per IBTimes UK. Stanley wrote that Grok is one of only four AI models cleared to support national security applications and one of three approved for mission-critical work in top-secret environments. The filing supports a Justice Department motion to dismiss an NAACP Clean Air Act suit over at least 57 unpermitted natural gas turbines that xAI runs at a Southaven, Mississippi, power plant to feed its $20 billion Colossus 2 data center in Memphis, Tennessee, per the Guardian. Associate Attorney General Stanley Woodward said "ultimate responsibility for enforcing federal law belongs to the Executive Branch, not private interest groups," per AP.

Read at The Guardian ↗ • Read at AP ↗ • Read at CNN ↗

IV.Industry & Market Watch

Intel rises 9% after Trump says Apple will partner on U.S. chipmaking

Intel shares jumped 9% in premarket trading Thursday after President Trump said on Truth Social that Apple has agreed to work with Intel to design and build chips in the U.S., per CNBC. Trump cast the arrangement as part of his domestic semiconductor rebuilding strategy. Neither company has publicly confirmed the scope, timing or commercial structure of the deal.

Read at CNBC ↗

Microsoft sells OpenAI models to Chinese tech giants through Azure as OpenAI and Anthropic stay out

Microsoft has built a large business selling OpenAI models to Chinese companies through Azure, even as OpenAI and Anthropic refuse to sell their own models in China over intellectual property and misuse concerns, Bloomberg reported. ByteDance has been its biggest AI customer in recent years and is on track to spend more than $1 billion a year on Microsoft AI and cloud services, with Ant Group, Meituan and Tencent also among major buyers. Azure's AI revenue grew faster in China than in any other territory, roughly tripling in the fiscal year ended June 2025. Under its OpenAI agreements, Microsoft does not host the models on Chinese soil over IP-theft fears; customers reach them from facilities abroad, such as Singapore.

Read at Bloomberg ↗

Anthropic joins Frontier carbon removal coalition as $915M tranche nearly doubles group's pledges to $1.8B

Anthropic became the first AI company to join Frontier, the corporate carbon removal buying group, contributing to a new $915 million purchase tranche, per the Wall Street Journal. The tranche brings Frontier's total pledged carbon removal contracting to about $1.8 billion since 2022, per TechCrunch. Frontier said it will narrow to 10 to 15 focused bets with 8- to 10-year purchase deals, and every new contract must show a "clear line of sight to government-driven demand" for the technology. The federal tax credit for CO2 removal currently applies only to direct air capture, leaving other removal methods reliant on corporate buyers like Frontier as the bridge to government demand, per Axios.

Read at WSJ ↗ • Read at TechCrunch ↗ • Read at Axios ↗

AI cyber warfare startup Twenty closes $100M Series B at $1B valuation for U.S. military and intelligence customers

Twenty Inc., founded in 2024 by former public sector cybersecurity officials, closed a $100 million Series B led by Accel at a $1 billion post-money valuation, per SiliconANGLE. Friends & Family Capital, Point72 Ventures and Caffeinated Capital joined the round. CEO Joe Lin, previously a vice president of product management at Palo Alto Networks, said the company is building the industrial base for American cyber power. Twenty has raised $138 million in total and will direct the new capital to research and development.

Read at SiliconANGLE ↗

V.Global & Geopolitics

Macron and Modi pitch G7 'trusted partners' guarantee on access to U.S. frontier AI

French President Emmanuel Macron told the G7 working lunch in Evian-les-Bains that the U.S. could "from one day to the next… turn off the switch" on allied access to American AI, per TechCrunch. Indian Prime Minister Narendra Modi, Canadian Prime Minister Mark Carney and Cohere CEO Aidan Gomez joined Macron in backing a multilateral "trusted partners" guarantee to lock in continued access. At his closing press conference Macron called Washington's Anthropic move "strictly nationalist" and said France will boost funding for its own AI industry as a hedge against a breakdown in cooperation, per the Washington Post. Macron paired the access pitch with a public call for binding regulation of frontier AI models to keep advanced systems away from authoritarian regimes, per Bloomberg.

Read at TechCrunch ↗ • Read at Washington Post ↗ • Read at Bloomberg ↗

UK Home Office moves ahead with facial age estimation for asylum seekers despite internal evidence of major errors

The UK Home Office plans to deploy facial age estimation AI on asylum seekers arriving at the border starting in 2027, despite an April 2025 internal report showing the systems regularly mistake children for adults and have substantial bias against Sub-Saharan Africans, per Wired. The report tested seven facial age estimation algorithms against more than 2.5 million images and found the best-performing system had an average error of 4.6 years on female Sub-Saharan Africans, meaning a 13.5-year-old could be classified as 18. The Home Office bought face scanning technology from German vendor Cognitec for more than $400,000 in May. Sixty-two rights groups including Foxglove sent an open letter Thursday asking the government to scrap the plans.

Read at Wired ↗

UK NCSC chief Horne reports 200+ critical infrastructure cyber incidents in past year, three-quarters tied to state actors

National Cyber Security Centre chief executive Richard Horne said the UK's critical national infrastructure absorbed more than 200 cyber incidents in the year to May, with about 75% believed to be linked to state actors including Russia, China and Iran, per the Guardian. Speaking at the Royal United Services Institute, Horne said advances in AI are likely to accelerate the threat and projected 2028 as the year an AI enabled threat to critical infrastructure crystallizes. He urged organizations to focus on cybersecurity fundamentals such as rapid recovery from attacks, warning that vulnerabilities tolerated in peacetime will be exploited in conflict.

Read at The Guardian ↗

EU Commission publishes 2026 State of the Digital Decade report, flagging AI uptake gap as priority for next MFF

The European Commission published its fourth State of the Digital Decade report Wednesday, identifying AI uptake and infrastructure as gaps standing between member states and the bloc's 2030 digital transformation targets. The report outlines priority reforms and investments meant to guide digital funding under the next EU Multiannual Financial Framework, the bloc's seven-year budget cycle. Henna Virkkunen, executive vice-president for Tech Sovereignty, Security and Democracy, said the bloc has adopted measures to strengthen Europe's capacity in semiconductors, AI, cloud and open source software. She said the EU must press on to achieve technological sovereignty. A Eurobarometer poll published alongside the report found a large majority of Europeans rank digital policy as a top EU priority.

Read at European Commission ↗