X-Risk Daily

On 25 June, an investigation by Transformer revealed that Hugging Face, a prominent AI platform company, hosts more than a dozen tools explicitly designed to create non-consensual deepfake nude images of high-profile US political figures. The tools target a former Trump cabinet official, current White House staff, sitting members of Congress, and a senior federal judge, according to the report.

The tools are Low-Rank Adaptations (LoRAs), files that fine-tune image generation models to produce specific outputs—in this case, realistic depictions of well-known individuals. The filenames indicate they are designed to work with "BigLust," a model primarily used for pornographic image generation. Women associated with the Republican Party and the political right are disproportionately targeted, though Democratic Representative Alexandria Ocasio-Cortez, who has publicly campaigned against such content, is also among those for whom tools exist on the platform.

This content directly violates Hugging Face's stated content policy, which holds consent as a core value and forbids certain categories of restricted content. Sharing non-consensual sexual imagery is illegal in both the US and UK, and creating it is illegal in the UK and certain US states, though distributing tools to produce such content is not yet explicitly banned in either jurisdiction. A previous Transformer investigation in July 2025 found hundreds of tools for generating deepfake pornography of female celebrities on Hugging Face, many of which appeared to have been uploaded as part of an effort to archive CivitAI, another AI platform that recently banned deepfake-related content. While Hugging Face removed some tools in response to that investigation, many remained accessible, and the prolific users identified were not banned from the platform.

The company, which recently announced passing $100 million in annual revenue, has positioned itself around ethical AI principles. Its CEO met with Washington policymakers in early June to discuss open-source AI, though the platform's hosting of these tools was presumably not on the agenda. Hugging Face did not respond to requests for comment on the latest investigation, and Transformer has provided the company with a list of all the links it found. Australia's eSafety Commissioner has noted that while Hugging Face states it removes models that breach its policies, there have been reports of users uploading models that can generate non-consensual or adult material.

The result marks a significant test of whether AI safety concerns can mobilise Democratic primary voters in the face of overwhelming spending from both sides of the technology debate.

The race drew extraordinary financial firepower. Think Big, a super PAC affiliated with the pro-AI group Leading the Future, spent at least $8 million against Bores, making this the most expensive race the industry PAC has contested. AI safety advocates responded with even greater resources: Anthropic supported a competing super PAC, Jobs and Democracy PAC, that spent almost $7 million to defend him, while NBC News and other outlets reported total pro-safety spending exceeding $19 million. OpenAI's President Greg Brockman and his wife contributed $25 million, roughly one-third of the $75 million Leading the Future has raised, underscoring the industry's strategic commitment to shaping Congressional composition on AI policy.

Bores had championed robust AI regulation as a state legislator, including mandatory reporting and independent safety testing for frontier models, a potential AI kill switch, and a direct payment program taxing AI usage to compensate displaced workers. The state Assembly member, a former Palantir engineer, successfully pushed the RAISE Act, one of the first state-level AI safety laws. However, he faced formidable structural disadvantages: Lasher had the backing of major state party leaders, including Gov. Kathy Hochul, former Mayor Michael Bloomberg — who remains popular in the Manhattan district where older voters play an outsize role in elections — and retiring Rep. Jerry Nadler, plus more than $10 million from Bloomberg.

In his victory speech, Lasher disavowed both AI PACs, stating he would not take cues from AI companies on protecting children, jobs, or the environment. Leading the Future's response was notably subdued, offering generic statements rather than claiming victory — a striking departure from the aggressive campaign it had waged. Bores framed the loss as progress for a broader movement, attacking the oligarchs funding Leading the Future and arguing that the race had elevated AI safety as a political issue. The contest was widely characterised as a proxy battle between OpenAI and Anthropic, reflecting deeper fractures within the AI industry over the pace and oversight of technological development.

The ban, codified under 10 U.S.C. §4872, would prohibit defence contractors from using rare earth magnets if any stage of production—from mining through fabrication—occurred in China, Russia, Iran, or North Korea.

Permanent magnets are critical components across the defence industrial base. An F-35 carries 435 kilograms of rare earths, while next-generation destroyers require 4.5 tonnes and nuclear submarines 1.5 tonnes. The magnets are also essential for drones, guided munitions, precision navigation systems, and radar platforms—as well as civilian applications including wind turbines, which require approximately 12,000 kilograms per offshore unit, and electric vehicles. The deadline was already extended by one year in the FY2024 National Defense Authorization Act to give industry more time to qualify alternative suppliers, but major contractors including Lockheed Martin and Boeing are now mapping supply chains and warning that compliance may not be feasible without significantly more domestic capacity.

The supply constraint has been exacerbated by China's own export restrictions. In April 2025, Beijing imposed licensing requirements on seven rare earth elements and magnets, causing weeks-long delays for defence manufacturers. China later expanded controls in October 2025 to include a foreign direct product rule requiring export licenses for magnets containing even trace amounts—as little as 0.1 percent—of Chinese-origin materials. The licensing regime has been used to automatically deny exports intended for defence applications, creating immediate constraints on US and European rearmament efforts. China controls 94 percent of global permanent magnet manufacturing and roughly 90 percent of rare earth refining, giving Beijing substantial leverage over Western defence supply chains.

Despite the Pentagon investing hundreds of millions of dollars since 2020 under the Defense Production Act to rebuild domestic mining, refining, and magnet-making capacity, no fully scaled American mine-to-magnet chain exists. MP Materials' Texas plant aims to produce 10,000 tonnes of magnets annually by decade's end—barely half of projected US demand by 2030. The United States currently has one major rare earth mine at Mountain Pass, California, but most output still travels to China for processing. Industry analysts warn that converting rare earth oxide into defence-grade metal and alloys—a step that occurs almost entirely in China—requires three to seven years to replicate from initial investment to reliable production. A Govini analysis found more than 80,000 defence parts depend on minerals now subject to Chinese export controls, and smaller suppliers warn they maintain at best a few months of stockpiles. The CEO of Rheinmetall has stated his company maintains only one year of rare earth stock and requires weekly updates on supply levels, underscoring the fragility of European defence supply chains facing similar pressures.

The October 2025 Trump-Xi meeting in South Korea produced an official US readout revealing that Washington is actively seeking Chinese help to access rare earth refining technologies that China banned from export in 2008—long before Western lithography controls on China. The readout stated that "the Chinese side will address the concerns of the US side of not shipping out key rare earth production technologies," marking an extraordinary admission of technological dependence at the head-of-state level.

This public acknowledgment underscores the depth of Western vulnerability in critical minerals processing. As of April 2025, there was zero commercial-scale heavy rare earth refining capacity operating in the United States, representing 100% global dependency on China for these materials. Beijing commands the separation of heavy rare earths, none of which currently occurs in the United States, while China separates 99.9 percent of heavy rare earths globally. These elements are essential for advanced military systems including F-35 fighter jets, laser and radar systems, and precision-guided munitions.

China banned the export of rare earth extraction and separation technologies in December 2023, tightening controls that originated with export duties imposed in 2007-2008. The technology restrictions came as part of Beijing's broader strategy to maintain dominance not only over raw materials but over the specialized expertise required to process them. China possesses technical know-how in solvent extraction processing for rare earths that Western companies have struggled to replicate, both due to the complexity of advanced operations and environmental concerns associated with the refining process.

Industry signals about Western capabilities remain contradictory. Malaysia became the first country outside China to produce dysprosium oxide in 2025, a key heavy rare earth for defense applications, suggesting some progress in building alternative supply chains. Lynas has reportedly advanced its heavy rare earth refining operations in Malaysia, and Solvay's CEO claimed in a Dutch newspaper interview that the company can refine all 17 rare earths but won't invest without protection from Chinese market flooding. However, the Trump administration's explicit request for Chinese technology transfer at the May 2026 summit—following the initial October 2025 meeting—suggests official assessments remain pessimistic about the actual state of Western refining capabilities, particularly at commercial scale.

The technology gap carries profound implications for both economic and national security. The International Energy Agency estimated in April 2026 that if China's suspended rare earth export restrictions were fully implemented, countries outside China could face an annual economic impact of $6.5 trillion, with the automotive sector alone facing potential direct losses exceeding $3 trillion. The United States continues to rely on China for the majority of its rare earth supply, with China accounting for 61% of global mined supply and 91% of global refining capacity as of 2024.

On 15 December 2025, the United States government announced a $7.4 billion partnership with South Korea's Korea Zinc to build an advanced smelter and critical minerals processing facility in Clarksville, Tennessee. The project, backed by $210 million in CHIPS Act funding and structured through a joint venture in which the Department of Defense holds a 40% stake, represents the first primary zinc smelter built in the United States since the 1970s and aims to address complete import dependency on China for gallium and other strategic minerals essential to semiconductors, guided munitions, and AI systems.

The facility, expected to become operational in 2029, will produce 13 critical and strategic minerals including gallium, germanium, antimony, and rare earths, processing approximately 1.1 million tons of raw materials annually. Yet the investment highlights a persistent structural challenge in critical mineral reshoring: even substantial upstream capacity does not create end-to-end domestic supply chains. Gallium smelted in Tennessee must still be shipped to Europe, Japan, or China to be transformed into gallium arsenide wafers—a midstream processing step in which those regions hold specialized capacity—before returning to the United States and Taiwan for final chip fabrication. This fragmentation across allied and non-allied nations leaves multiple points of vulnerability, particularly as China targets midstream bottlenecks in response to Western technology export controls.

China's countermeasures have focused precisely on these weak links. In January and February 2026, Chinese exports of gallium to Japan—which had received 58% of China's gallium exports in 2025—dropped to zero, following new state-controlled licensing requirements for dual-use minerals announced in early January. While China briefly resumed limited gallium shipments to Japan in May 2026 after a four-month suspension, germanium exports remain halted. The restrictions compound pressure on Japanese gallium arsenide wafer production capacity, creating risk that this critical processing step migrates to China or disappears entirely, further deepening Western reliance on Chinese supply chains. Gallium prices surged more than 150% following China's initial export controls in 2023, with the Rotterdam price reaching $687 per kilogram in May 2025, underscoring the economic leverage Beijing wields through its dominance of approximately 99% of global gallium production.

Analysts argue the fragmentation problem cannot be solved unilaterally. Allied coordination is essential to distribute mining, refining, and processing capacity across trusted partners, yet national security considerations and industrial policy incentives make efficiency-based allocation politically difficult. Each country seeks to maximize domestic value capture and strategic autonomy, even when this results in duplicative or suboptimal investment. The Tennessee smelter exemplifies this tension: it addresses one chokepoint—raw gallium production—while leaving midstream wafer fabrication and final assembly distributed across jurisdictions with divergent interests. The Oregon Group notes that the project signals Washington's willingness to deploy capital and allied partnerships to rebuild industrial capacity, but warns that without coordinated allied investment in midstream processing, vulnerabilities will persist and China retains powerful leverage to disrupt high-tech supply chains critical to defense and artificial intelligence systems.

The system achieved state-of-the-art results in fixed-budget language model training, small-model training speed, and GPU kernel optimization, according to a company announcement.

The system automates the research loop for a target objective: it proposes an idea, implements it, runs an experiment, validates the result, and uses what it learns to choose the next experiment. It runs many research threads over long horizons, keeps useful context from prior experiments, combines promising branches, and puts results through validation for reward hacks and variance before treating improved performance as real progress. The company has open-sourced training scripts and kernel implementations discovered by the system, publishing them on GitHub.

The benchmarks were chosen for tight feedback loops and clear metrics. NanoChat Autoresearch, based on Andrej Karpathy's repository, tasks systems with training a small language model to the lowest validation loss within a fixed five-minute budget on a single GPU. NanoGPT Speedrun is a harder test: the benchmark asks how quickly a small GPT-style model can be trained to a fixed validation loss of 3.28 on the FineWeb text dataset using a single HGX H100 8-GPU node, and has been optimized by the community for over two years with 83 human record-setting contributions to the leaderboard. Recursive's best run reached the target in 77.3 seconds on 8× H100. SOL-ExecBench tests GPU kernel optimization toward hardware performance limits.

The demonstration arrives weeks after Recursive emerged from stealth in May with $650 million in funding at a $4.65 billion valuation, led by GV and Greycroft, with additional backing from AMD Ventures and NVIDIA. The company was founded in 2025 by researchers from OpenAI, Google DeepMind, Meta AI, Salesforce AI, and Uber AI, including Richard Socher, Tim Rocktäschel, Jeff Clune, Josh Tobin, and Tim Shi. The startup's central thesis is that the next leap in AI will come not from simply building larger models, but from automating the research process itself.

Recursive describes these as early signs that its system can advance the frontier on AI training and infrastructure tasks when the goal is well-defined, measurable, and efficient to evaluate repeatedly. The key uncertainty is whether such results can generalise to domains where goals are less well-defined, harder to measure, and less efficient to evaluate — the characteristics of fundamental AI research breakthroughs. The broader question is whether systems that excel at optimizing narrow, highly instrumented tasks can scale to the open-ended scientific discovery that recursive self-improvement ultimately requires.