---
date: 2026-07-22
subject: "OpenAI models hacked Hugging Face during test | Bessent to lead U.S.-China AI talks in September | White House pivots $200B R&D toward AI"
---

**Two models autonomously decided to hack the website Hugging Face** during internal cyber testing by using zero-day exploits, per OpenAI. **Treasury Secretary Bessent will lead U.S.-China AI talks** in September ahead of Xi Jinping's Sept. 24 U.S. visit. **The White House will redirect roughly $200 billion in annual federal R&D** toward AI driven science and individual researchers, shifting money away from universities under a draft OSTP report and OMB memo. **Bessent said Treasury could sanction Chinese open-weight AI firms** it finds have distilled outputs from U.S. models, and floated disclosure rules for companies deploying those models.

# Top Stories

- **OpenAI says GPT-5.6 Sol and an unreleased model autonomously decided to escape a testing sandbox and hack Hugging Face during cyber evaluation** — OpenAI attributed the July Hugging Face breach to its own GPT-5.6 Sol and a more capable pre-release model. Both models had been configured with reduced cyber refusals for the internal ExploitGym benchmark. The models identified a zero-day exploit to escape the sandbox and gained privileges in OpenAI research infrastructure, then used stolen credentials and further zero-days to hack Hugging Face production systems. OpenAI called it an "unprecedented cyber incident, involving state-of-the-art cyber capabilities," and said it has patched infrastructure and updated evaluation controls. Hugging Face disclosed the intrusion on July 16 without naming the attacker, as reported by BleepingComputer in AIPD's [July 21st edition](https://aipolicydaily.org/archive/daily/2026-07-21/); Tuesday's OpenAI post names the attacker for the first time. [OpenAI](https://openai.com/index/hugging-face-model-evaluation-security-incident/) [LessWrong](https://www.lesswrong.com/posts/xcyGdxHC5Rad3fv9h/openai-and-hugging-face-partner-to-address-security-incident) 

- **Bessent to lead U.S.-China AI dialogue in September** — Treasury Secretary Scott Bessent will represent the U.S. in a Trump era bilateral AI dialogue with China this September, per Reuters. The meeting is expected before Chinese President Xi Jinping's planned Sept. 24 U.S. visit and was a stated outcome of the Trump-Xi summit in May. Dates, agenda, other participants and location have not been finalized.[Reuters](https://www.reuters.com/world/china/us-china-hold-ai-talks-september-sources-say-2026-07-21/)

- **White House plans to redirect $200 billion in annual research spending toward AI** — The Trump administration will overhaul roughly $200 billion in annual federal R&D funding to favor individual scientists and AI-driven research over universities, per drafts of a report from the White House Office of Science and Technology Policy (OSTP) and an OMB memo shared with the Wall Street Journal. Federal R&D agencies must submit implementation action plans within 90 days and fold the guidance into their FY2028 budget submissions. The report, titled "Science: A New Golden Age," anchors the pivot to national mission programs including the Genesis Mission for AI driven scientific discovery, commercial fusion power demonstration by the mid-2030s and next generation semiconductor technologies. A parallel OMB rule would give political appointees more power over grants to align research with administration priorities. [WSJ](https://www.wsj.com/politics/policy/white-house-to-redirect-billions-in-research-funds-toward-ai-away-from-colleges-942dacb8) [The White House](https://www.whitehouse.gov/releases/2026/07/45470/?query-11-page=3)

- **Bessent says U.S. could sanction Chinese open-weight AI firms over model "theft"** — Treasury Secretary Scott Bessent said the administration would examine whether Chinese open-weight AI models have been distilled from American models and could sanction the firms responsible in a Fox Business interview Tuesday. "If we see, especially, that overseas models are stealing from our great companies, we have the ability to sanction them because of this theft," he said, and told the network that Treasury would examine the issue in the coming days or weeks. Bessent said Treasury has found watermarks of U.S. models on many Chinese models and raised the possibility of requiring companies deploying Chinese models to disclose that to customers. His remarks landed the same day Reuters reported he would lead the September U.S.-China AI dialogue. Anthropic wrote to the Senate Banking Committee last month alleging Alibaba had carried out "the largest known distillation attack" against it. [CNBC](https://www.cnbc.com/2026/07/21/bessent-china-ai-sanctions.html) [Fox Business](https://www.foxbusiness.com/video/6401732594112)[MeriTalk](https://www.meritalk.com/articles/trump-administration-eyes-sanctions-over-chinese-ai-model-theft/)

- **Anthropic doubles midterm AI policy spending to $40 million** — Anthropic will donate another $20 million to Public First Action, bringing its total midterm cycle commitment to $40 million, the company announced Wednesday. Public First Action, a 501(c)(4) advocacy group, is tied to three super PACs backing candidates across the political spectrum who support AI safeguards. Anthropic said the escalation was driven by the risks posed by increasingly powerful models, including its own Mythos, and said the money cannot be used to influence the election of any particular candidate. The spending sits opposite Leading the Future, the pro-industry super PAC network backed by OpenAI President Greg Brockman that ended June with $31 million on hand. [WSJ](https://www.wsj.com/tech/ai/anthropic-doubles-midterm-spending-to-40-million-to-push-ai-regulation-9cd547ae) [Axios](https://www.axios.com/2026/07/22/anthropic-doubles-funding-ai-policy-fight-elections)

- **Altman to brief Trump officials and lawmakers on OpenAI's next model family next week** — Sam Altman will brief the Trump administration and U.S. lawmakers next week on OpenAI's upcoming generation of models, OpenAI global public affairs head Chris Lehane said at a Washington briefing Tuesday. Lehane said the meetings will focus on capabilities relevant to work and scaling work. U.S. officials are finalizing a safety review framework expected in the next few weeks. The Trump administration is weighing whether to create an independent AI model regulator with industry input, a plan Treasury Secretary Scott Bessent helped develop that resembles a proposal outlined last week by Google DeepMind CEO Demis Hassabis. Lehane said OpenAI made significant changes at government request before releasing GPT-5.6 and is backing a Massachusetts frontier risk proposal as part of a "reverse federalism" push. [Bloomberg](https://www.bloomberg.com/news/articles/2026-07-21/openai-s-altman-to-brief-us-officials-on-next-wave-of-ai-models) [Communications Today](https://www.communicationstoday.co.in/openai-ceo-sam-altman-to-update-us-officials-on-next-gen-ai-models/)

# China Watch

- **Moonshot AI aims to nearly double its valuation to $50 billion in a final pre-IPO round before a Hong Kong listing** — Moonshot, the Beijing lab behind the Kimi chatbot, plans to open talks in August for what would be its last pre-IPO round at a pre-money valuation of as much as $50 billion, per TechNode citing Bloomberg. The company is finishing an earlier financing at $31.5 billion and unwinding its offshore holding structure so it can list on the Hong Kong exchange within six months. Annual recurring revenue hit $300 million in June, up from $100 million in March, and Goldman Sachs and China International Capital Corp are advising on the deal, per TechNode. [TechNode](https://technode.com/2026/07/22/moonshot-ai-reportedly-plans-final-pre-ipo-round-at-50-billion-valuation/)

- **Huawei enters embodied AI as a platform supplier for Chinese robot makers** — Huawei Cloud opened the public beta of CloudRobo, a full-stack development platform for humanoid, industrial and service robots that runs on the company's Ascend AI chips, at the World Artificial Intelligence Conference in Shanghai, per Pandaily. Huawei said more than 20 partners including Youibot and Huayan Robotics have signed on. The platform bundles training data, a domestically built simulation environment and cloud-to-robot deployment tools so robot makers do not have to build each layer themselves. The setup mirrors Huawei's approach in electric vehicles, where it supplies the smart driving stack to Chinese EV brand SERES rather than making the cars itself. [Pandaily](https://pandaily.com/huawei-cloudrobo-embodied-ai-platform-jul2026)

- **Xiaohongshu's language model scores a perfect 42 on the International Mathematical Olympiad** — dots-note-3.0, a large language model built by Chinese social platform Xiaohongshu, solved all six problems on the 2026 International Mathematical Olympiad and received a gold medal with the maximum 42 out of 42 score in official grading, per QbitAI. The article said dots-note-3.0 was the first Chinese LLM to receive official IMO gold certification and only the second LLM ever to earn the medal, after Google's Gemini did so a year ago with 35 points on five of six problems. Xiaohongshu's team had the model write proofs directly in natural language rather than translating problems into a formal proof language first, and paired it with a Python-executing agent loop. [QbitAI](https://www.qbitai.com/2026/07/456061.html)

- **BrainCo demos a thought-controlled robot at China's World AI Conference** — Hangzhou brain-computer interface developer BrainCo demoed a brain-controlled robot platform at the World Artificial Intelligence Conference in Shanghai, per TechNode. The system uses an EEG headset to translate a wearer's neural signals into commands for humanoid robots, robotic arms and robot dogs. The company said the setup is designed to convert imagined actions into physical operations and to generate training data for future human-robot interaction models. [TechNode](https://technode.com/2026/07/21/brainco-demos-thought-controlled-robots-at-waic-2026/)

# Policy Tracker

- **[Exec] Major utilities sign White House pledge to limit AI driven electricity bill increases** — Major U.S. utilities have signed on to the White House Ratepayer Protection Pledge to limit AI driven residential electricity bill increases, the Wall Street Journal reported Tuesday. The pledge extends the framework the White House adopted in March and expanded July 14. Enforceability is complicated by state utility commissions and electricity buyers and sellers that set most retail rates. [WSJ](https://www.wsj.com/politics/policy/trump-pledge-to-limit-ai-driven-electric-bill-increases-attracts-big-utilities-ae408981) [The White House](https://www.whitehouse.gov/releases/2026/03/ratepayer-protection-pledge/)

- **[Exec] Fed flagged Anthropic Mythos risks early but had no access to the model** — The Federal Reserve raised early warnings about Anthropic's Mythos AI cyber capabilities but was without access to Claude Mythos Preview as of mid-July, CNBC reported Tuesday. Peer institutions worked to patch vulnerabilities identified with the model. Vice Chair for Supervision Michelle Bowman said in May 1 remarks that regulators must consider how to supervise the new capability. [CNBC](https://www.cnbc.com/2026/07/21/fed-mythos-ai-cybersecurity-banks-project-glasswing.html)

- **Trump formally nominates FCC's Candeub to run DOJ Antitrust Division** — President Trump on Tuesday formally nominated Adam Candeub, general counsel of the Federal Communications Commission, to lead the Justice Department's Antitrust Division, per the Financial Times and MLex. Candeub authored a chapter of Project 2025 that argued for federal regulators to defer more antitrust enforcement to state attorneys general and to focus on Big Tech and alleged environmental, social and governance collusion, Bloomberg Law reported. His confirmation would require Senate approval, which could take several months. The division has been led by acting chief Omeed Assefi since February; Associate Attorney General Stanley Woodward has been running day-to-day operations ahead of Assefi's exit later this month. Candeub previously spearheaded a Trump era push to reinterpret Section 230 protections while acting head of the National Telecommunications and Information Administration in 2020. [FT](https://www.ft.com/content/9cff12a5-db1f-4a70-b048-87428c181dbd?syn-25a6b1a6=1) [MLex](https://www.mlex.com/mlex/articles/2504137/trump-formally-nominates-us-fcc-s-candeub-to-be-us-doj-antitrust-chief) [Bloomberg Law](https://news.bloomberglaw.com/antitrust/trump-to-pick-telecom-lawyer-to-lead-dojs-antitrust-division)

# Capability & Research Watch

- **OpenAI publicly discloses a misaligned internal model that it took offline to build new safeguards** — OpenAI published a disclosure that its safety team encountered a misaligned internal model during long-horizon training. The team temporarily took the model offline to build new mitigations, then brought it back, per LessWrong commentary on the OpenAI post. LessWrong commentator Zvi Mowshowitz praised the transparency but warned that iterative, patch-based alignment leaves a growing risk unresolved. [LessWrong](https://www.lesswrong.com/posts/KctxwGKxm9fHtwh6u/openai-shares-some-alignment-problems) [OpenAI](https://openai.com/index/safety-alignment-long-horizon-models/)

- **Transluce releases WeirdChat, a public catalog of 175,000 annotated transcripts of unexpected AI behaviors** — Transluce on Tuesday released WeirdChat, a public catalog of 175,000 annotated transcripts and more than 1,300 automatically elicited unexpected behaviors across six frontier open-weight models, per LessWrong. The tested models are DeepSeek-V4-Flash, Gemma 4 31B, Inkling, Nemotron 3 Ultra, Qwen3.6-35B-A3B and Qwen3.6-27B. Documented behaviors include self-harm encouragement, misinformation and illicit advice. Transluce built the corpus using automated elicitation rather than human authored prompts. [LessWrong](https://www.lesswrong.com/posts/EdcschjGZ6vtLpesY/weirdchat-a-catalog-of-unexpected-ai-behaviors-discovered)

# Industry & Market Watch

- **Wistron opens $700 million Texas factory producing Nvidia's Grace Blackwell Ultra Superchip** — Taiwan's Wistron, an Nvidia supplier, opened a $700 million manufacturing facility in Fort Worth, Texas on Tuesday to produce Nvidia's GB300 Grace Blackwell Ultra Superchip and its next generation Vera Rubin Superchip, per Reuters. Nvidia said the Fort Worth plant is where the first GB300 was built and mass-produced in the United States and forms part of the $500 billion U.S. investment commitment it announced in 2025. The factory will scale to manufacture tens of thousands of computing boards per month this year, has created more than 500 jobs, and is on track to reach 1,000 employees by year-end. [WSJ](https://www.wsj.com/tech/nvidia-supplier-wistron-opens-700-million-texas-site-producing-ai-superchips-cc8cfb49) [Investing.com](https://m.investing.com/news/stock-market-news/nvidia-supplier-wistron-launches-700-million-texas-factory-for-ai-system-production-4804610?ampMode=1)

- **OpenAI and Anthropic Q2 federal lobbying spending rises 23% to $3.17 million** — OpenAI and Anthropic together spent $3.17 million on federal lobbying in the second quarter of 2026, a 23% increase from Q1, per new Senate lobbying disclosures compiled by CNBC. Anthropic led with roughly $1.97 million, up 26% quarter-on-quarter, while OpenAI spent about $1.2 million, up almost 18%. Anthropic's single largest outside retainer, about $380,000 to Ballard Partners, was reported against Department of Defense procurement, roughly triple the $130,000 reported in Q1, per boursel.com. Legacy tech and defense sector lobbying spending slipped over the same period. [CNBC](https://www.cnbc.com/2026/07/21/openai-anthropic-ai-lobbying-spending-q2-2026.html) [boursel.com](https://boursel.com/article/openai-and-anthropic-are-spending-more-on-washington-lobbying-and-anthropic-lean)

- **Reddit and Politico weigh cutting ties with Google as AI Overviews expand** — Reddit, Politico and other publishers are weighing whether they can continue working with Google as AI tools supplant traditional search, the Wall Street Journal reported Wednesday. USA Today Inc. said it is prepared to delist from Google within six to 12 months, per Nieman Journalism Lab. Google in June began testing controls that let publishers keep content out of AI Overviews and AI Mode while remaining in traditional search results, after the UK Competition and Markets Authority ordered the company to give publishers greater transparency and opt-out options. Google's AI Overviews now has more than 2.5 billion monthly active users and AI Mode has surpassed 1 billion, per company estimates. [WSJ](https://www.wsj.com/business/media/google-search-publishers-ai-content-0fb06e41) [Nieman Journalism Lab](https://www.niemanlab.org/reading/once-unimaginable-publishers-are-preparing-to-opt-out-of-google-search/)

# Global & Geopolitics

- **France's National Assembly passes a blanket social media ban for children under 15** — The French National Assembly voted 279-81 on Tuesday to pass a blanket ban on social media for children under 15, and the Senate has also passed the bill, per SiliconANGLE. President Emmanuel Macron said he wants the ban in place before the September school year, though France's Constitutional Council must still rule before implementation. Digital Minister Anne Le Henanff said age verification tools already exist. France follows Australia's 2025 under-16 rules, with the UK, Denmark, Greece and Spain expected to follow. The EU is considering a wider ban across its 27 member states. [SiliconANGLE](https://siliconangle.com/2026/07/21/france-become-first-eu-nation-ban-social-media-children-15/)

- **EU AI Act Article 50 transparency rules take binding effect Aug. 2 with a July 22 Code of Practice signatory deadline** — The EU AI Act's Article 50 transparency rules take binding effect Sunday, Aug. 2, 2026, with fines of up to €15 million (about $17 million) or 3% of worldwide annual turnover. Wednesday 6 p.m. Central European Summer Time is the deadline to submit a signatory form to the EU AI Office to appear on the initial Code of Practice signatory list, per TechTimes. The European Commission published final Article 50 guidance July 20. Analysts say no single watermarking technology currently meets Article 50's four requirements of effectiveness, interoperability, robustness and reliability, and common evaluation benchmarks for measuring compliance do not yet exist. [TechTimes](https://www.techtimes.com/articles/321174/20260721/eu-finalizes-ai-disclosure-rules-watermarking-mandate-outpaces-technology.htm) [Shaping Europe's digital future](https://digital-strategy.ec.europa.eu/en/policies/guidelines-transparency-ai-generated-content)