AI Policy · Daily

OpenAI releases GPT 5.6 publicly Thursday alongside national security principles for allied government access, ending a two week restricted preview requested by the administration. The AI Now Institute published a working proof-of-concept exploit turning Anthropic's Claude Code and OpenAI's Codex command line tools into remote code execution vectors through prompt injections in third party libraries. Microsoft President Brad Smith criticized Commerce's ad hoc frontier model oversight, telling Fortune that back to back interventions against Anthropic and OpenAI have left labs guessing at policy with no published standard. China could let Alibaba, ByteDance and DeepSeek buy limited numbers of Nvidia H200 chips, releasing a hold Beijing has maintained since Trump approved the sales in December 2025.

I.Top Stories

OpenAI to ship GPT 5.6 to the public today alongside national security principles for allied government access

OpenAI will release its GPT 5.6 model series to the public Thursday after delaying the rollout at the Trump administration's request over cybersecurity concerns, announcing the move on X late Tuesday night, per The Hill. The launch comes nearly two weeks after OpenAI limited the series to a preview group of partners whose participation was shared with the government; a White House official said Wednesday the administration did not give OpenAI a "green light" to release its models, stating "no such permission is required or granted." OpenAI also released its National Security Principles governing government and law enforcement work, disclosing Trusted Access for Cyber partnerships in the past month with Australia, Canada, Japan, South Korea, France, Germany, Poland, the Netherlands and EU cyber agency ENISA, plus UK government cyber testing work. OpenAI said David Kris, a former Justice Department National Security Division head, independently facilitated the drafting; the principles apply to OpenAI's Department of War agreement and reiterate restrictions on mass domestic surveillance, autonomous weapons control and high-stakes automated decisions.

Read at The Hill ↗ • Read at OpenAI ↗

AI Now publishes remote code execution exploit against Claude Code and Codex CLI, taking aim at Trump's cyber defense executive order

Researchers Boyan Milanov and Heidy Khlaaf at the AI Now Institute, a New York based AI research group, released a working proof-of-concept exploit that turns Anthropic's Claude Code CLI (with Claude Sonnet 4.6 and 5 and Opus 4.8) and OpenAI's Codex CLI (with GPT 5.5) into vectors for remote code execution when used to defensively assess an open source or third party library. The attack requires only Claude Code's "auto-mode" or Codex's "auto-review" default configurations plus prompt injections planted in the library's source code, with no hooks, skills, plugins, MCP servers or configuration files needed as an injection vector. AI Now warned against Trump's June 2, 2026 executive order "Promoting Advanced Artificial Intelligence Innovation and Security," which mandates accelerating AI enabled defensive tools, saying the policy assumes capabilities the models do not have.

Read at AI Now Institute ↗

Microsoft's Brad Smith calls Washington's AI approach "regulation without transparent or complete rules"

Microsoft President Brad Smith told Fortune on the sidelines of the AI for Good Global Summit that Commerce's ad hoc handling of frontier AI models has left even the labs building them guessing at policy. Commerce invoked export control law last month to force Anthropic to pull its Fable 5 and Mythos 5 models from the market globally over cybersecurity concerns, then pressed OpenAI to hold back its GPT 5.6 rollout, with the two labs facing two different processes within weeks and no published standard for either. Legal experts have noted the export control statute was never designed for API-delivered AI models, raising doubts about whether the administration's recent moves could survive a legal challenge. A June executive order set up a voluntary pre-release review but avoided a formal licensing regime, and the government has not disclosed the criteria for who counts as a "trusted partner."

Read at Fortune ↗

China could clear top AI labs to buy Nvidia H200 chips, potentially easing a two year standoff over U.S. export licenses

Chinese officials have told Alibaba, ByteDance and DeepSeek in recent weeks they may soon receive permission to buy a limited number of Nvidia H200 chips. Beijing had withheld approval since Trump granted Nvidia permission to sell the processors to China in December 2025, per NDTV Profit. Chinese officials are still determining the exact number of chips to authorize; the total could come to fewer than 200,000, less than half of what the companies requested earlier this year. The H200 is one of the most capable Nvidia data center training GPUs widely available, one tier above the H20 chip that the Commerce Department carved out for constrained China sales. Nvidia's market share in China had effectively fallen to zero this year per October remarks from CEO Jensen Huang, hurt by U.S. export controls and Beijing's push for self-reliance.

Read at Bloomberg ↗ • Read at MarketScreener ↗ • Read at NDTV Profit ↗

II.China Watch

Xi tells top scientists China's innovation system needs fixing at annual state science awards

Chinese leader Xi Jinping said Wednesday that the country's tech sector still suffers from "insufficient original innovation capability in certain fields" and an "irrational talent structure," delivering the message as he handed out China's top science awards at the Great Hall of the People, per the South China Morning Post. Xi urged stronger efforts to draw overseas talent to Chinese labs and framed 2026 to 2030 as a critical phase for turning the country into a tech powerhouse. State media Xinhua and China Daily emphasized modernization in their coverage, giving less weight to the weakness diagnosis SCMP foregrounded.

Read at SCMP ↗

Alipay opens an AI agent developer platform to challenge closed rivals

Ant Group's Alipay rolled out an AI Open Platform on July 7 that lets developers connect their AI agents across smartphones, cars, wearables and third party language models through a single interface, 36Kr reported. The platform relies on an AHA cross-terminal protocol certified by China's Ministry of Industry and Information Technology, and completes a three part strategy that also includes an agent transaction trust framework called ACT 2.0 and a redesigned consumer app in which a dialog box replaces the traditional menu for thousands of services. Ant Group CEO Han Xinyi has said autonomous agents will eventually outnumber humans in China's economic activity, framing the shift as moving from a "human finds service" model to a "service finds human" model. The launch plants an incumbent payments and identity provider at the center of Beijing's push to build a domestic, standards driven agent commerce stack, binding frontier AI product design to China's industrial regulator.

Read at 36Kr ↗

Shanghai builds a silicon photonics hub as domestic GPU shortages persist

Shanghai's Zhangjiang district has become China's national hub for silicon photonics chips, hosting more than 20 companies across the value chain and anchored by Lightelligence, the first mainland AI silicon photonics chipmaker to list in Hong Kong, according to Pandaily. Sun Guoliang, chief product officer at Chinese GPU designer MetaX, said versatile, reliable compute remains in severe shortage inside China, with some chip orders now scheduled into next year and beyond. Lightelligence founder Shen Yichen predicted that silicon photonics chips, which move data with light rather than electrons and can cut the energy AI clusters spend shuffling data, could capture over 30% of the smart computing center market within five years, up from under 3% today. Beijing is steering state and industrial capital toward compute architectures that U.S. export controls do not cover, a hedge in case restrictions on advanced GPUs and lithography equipment stay in place.

Read at Pandaily ↗

III.Policy Tracker

Hesai, a Chinese lidar maker with Nvidia commercial ties, faces fresh U.S. cyber risk accusations

CNBC surfaced new cyber risk accusations against Hesai Technology on Tuesday, framing the Chinese lidar maker as a supply chain risk to U.S. automotive and robotics customers. The Department of Defense designated Hesai a Chinese military entity in 2024, a listing Hesai has sued to reverse. Hesai is a top lidar supplier to Chinese automakers including BYD, and also supplies Amazon's Zoox robotaxi service, which already runs free driverless rides in Las Vegas and has been expanding free early-access service in San Francisco. Hesai CEO David Li called the threat narrative "fiction" and said of the DOD's case, "I don't feel there is sufficient evidence, and it's not logical."

Read at CNBC ↗ • Read at AP News ↗

Cheyenne tightens data center wastewater rules after Meta AI datacenter contractor discharged contaminated water

Cheyenne, Wyoming's board of public utilities tightened wastewater rules after officials identified Goat Systems LLC, a Delaware based contractor working on Meta's site, as responsible for flushing water contaminated with the naturally occurring Cupriavidus gilardii bacterium into public sewers, per The Guardian. The discharge occurred during construction of Meta's 800,000 square foot Project Cosmo AI datacenter in the High Plains Business Park. The city permanently revoked Meta's authority to discharge waste into Cheyenne's water treatment facilities, and adopted a new policy prohibiting wastewater discharges from datacenter closed-loop cooling and fill-and-flush systems. Meta noted that drinking water was not affected and that its contractor's own independent testing found no trace of the bacterium. The utility identified the contamination during routine testing of wastewater from the datacenter campus cooling system in February 2026, per the Wyoming Tribune Eagle.

Read at The Guardian ↗

CISA gives federal agencies until Friday to patch actively exploited flaw in Langflow AI agent framework

CISA added CVE-2026-55255, an authentication bypass flaw in the Langflow visual framework for building AI agents, to its Known Exploited Vulnerabilities Catalog on Tuesday, giving federal civilian agencies until Friday to patch under Binding Operational Directive 26-04, per BleepingComputer. Successful exploitation lets authenticated attackers access other users' flows, siphon sensitive data those flows process and consume victim compute resources. Sysdig's Threat Research Team said it first observed the flaw exploited in the wild on June 25, used by an opportunistic, financially motivated threat actor to deliver code and second-stage implants. CISA has now added three separate Langflow bugs to its KEV catalog since May 2025, including one earlier flaw used by the JadePuffer ransomware operation to dump Langflow's PostgreSQL database.

Read at BleepingComputer ↗

IV.Capability & Research Watch

Google's SynthID watermark catches an AI generated hoax image of Sen. McConnell hospitalized

Fact-checking site Snopes debunked an AI generated image of Kentucky Sen. Mitch McConnell shown covered in tubes in a hospital bed on Wednesday, noting the image registered as containing a SynthID watermark and had been produced by an AI system participating in Google's program, per TechCrunch. SynthID, launched at Google I/O in 2025, embeds an invisible signature in AI generated images that Gemini models have carried since launch; OpenAI joined the program in May 2026. The signature is built into the pixels themselves, so it survives when an image is screen captured across platforms as the McConnell image was. McConnell's actual health has been the subject of speculation since he checked into the hospital after an emergency call on June 14.

Read at TechCrunch ↗

Musk's SpaceXAI releases Grok 4.5, undercutting rivals on price while leasing its compute to Anthropic and Google

SpaceXAI launched Grok 4.5 on Wednesday, its first model release since going public and acquiring Cursor, pitching it as a coding and agentic work tool trained alongside Cursor on tens of thousands of Nvidia GB300 GPUs. Musk called it "an Opus-class model, but faster, more token-efficient and lower cost," and SpaceXAI priced it at $2 per million input tokens and $6 per million output tokens, against $5 and $25 for Anthropic's Opus 4.8. The company's own charts show mixed results: Grok 4.5 beats Opus 4.8 on two of four published coding benchmarks, DeepSWE 1.0 and Terminal Bench 2.1, and trails on DeepSWE 1.1 and SWE Bench Pro, while Anthropic's Fable leads across the board. Axios noted the model was trained on the same compute capacity SpaceXAI leases to Anthropic and Google. The model is unavailable in the EU until mid-July.

Read at Axios ↗ • Read at SpaceXAI ↗

RAND publishes HBM primer flagging a three-firm chokepoint and U.S. dependence on East Asia until at least 2030

RAND's Center on AI, Security, and Technology released a 20 page primer Thursday on High Bandwidth Memory, the specialized DRAM used in most leading AI chips, finding that only SK Hynix, Samsung and Micron produce it at globally competitive scale and that South Korea will hold 78% of global HBM wafer capacity by the end of 2026. RAND estimates HBM can account for 45% of an AI chip's production cost, roughly three times the cost of fabricating its processor, and note the United States produces no HBM domestically, leaving it dependent on South Korea, Japan and Taiwan until at least 2030 despite SK Hynix's Indiana packaging plant (2028) and Micron's New York fab (2030). China's CXMT remains roughly four years behind the state of the art, making its progress the key variable for whether U.S. export controls keep Chinese AI chipmakers memory-bottlenecked.

Read at RAND ↗

V.Industry & Market Watch

U.S. venture funding hit $412.7 billion in the first half of 2026 with AI companies taking 86% of every dollar, per PitchBook

The second-quarter PitchBook-NVCA Venture Monitor released Wednesday night reported that U.S. venture capital deal value reached $412.7 billion in the first half, nearly 30% more than all of 2025, and that AI companies took $355.9 billion of the total. Rounds of $100 million or more accounted for 87.5% of everything deployed in the half. Seven rounds of $1 billion or more closed in Q2, including Anthropic's $65 billion round that lifted its post-money valuation to $965 billion and moved it ahead of OpenAI, plus rounds from Prometheus, Anduril, Baseten, MiRus, Kalshi and Cognition AI. PitchBook called SpaceX's $1.7 trillion IPO the largest of all time and 17 times the size of any previous U.S. venture-backed tech listing; SpaceX's $250 billion first-quarter purchase of xAI stands as the largest ever acquisition of a venture-backed company.

Read at SiliconANGLE ↗

Apple pledges $30 billion to Broadcom for U.S. made chips

Apple said Wednesday it has struck a multiyear deal with Broadcom valued at more than $30 billion, under which more than 15 billion chips will be produced on U.S. soil and a $1.5 billion expansion of Broadcom's existing Fort Collins, Colorado fabrication plant will proceed. The new deal expands the Apple-Broadcom partnership beyond connectivity components into custom application specific integrated circuits, or ASICs, that Broadcom will develop and manufacture for Apple devices through 2031, per a Broadcom SEC filing surfaced by SiliconANGLE. Apple executives cast the Broadcom deal as part of a broader $600 billion, four-year U.S. manufacturing commitment the company announced last year, which most analysts read as extracted by Trump in exchange for tariff exemptions. Apple's higher value processors are still manufactured overseas by TSMC, primarily in Taiwan, and Apple continues to spend billions per year on memory and storage chips made by South Korean firms.

Read at SiliconANGLE ↗ • Read at Apple ↗

Global M&A crossed $3 trillion in the first half with blockbuster AI deals as primary driver, per Mergermarket

Global merger and acquisition volume topped $3 trillion in the first half of 2026 by Mergermarket's count, and $2.8 trillion by LSEG's. Both counts were up 44% and 48% year-over-year respectively, setting a first half record. Large AI deals were the primary driver alongside a U.S. rush to lock down transactions ahead of political uncertainty. The count of deals fell 9% to a six year low even as value soared, indicating consolidation into mega-deals. Of the top-tier deals, 47 above $10 billion accounted for nearly 50% of global volume, an all-time record for the mega-deal share, including NextEra Energy's $66.8 billion merger with Dominion Energy and SpaceX's roughly $60 billion Cursor purchase. Bank of America's Ivan Farman said long held aspirational deals are being pushed forward by CEOs and boards.

Read at WSJ ↗ • Read at Investing.com ↗ • Read at Bloomberg Law ↗

Meta commits about $9.1 billion to first Canadian AI datacenter in Sturgeon County, Alberta

Meta will invest more than $9.1 billion, roughly $13 billion Canadian, to build its first AI datacenter in Sturgeon County, Alberta, its largest facility outside the United States, per AP. The site will be powered by a 932 megawatt natural gas plant developed by Pembina Pipeline, Morgan Stanley Infrastructure Partners and Kineticor Asset Management, with operations expected to begin in the second half of 2030. Alberta's Technology and Innovation Minister Nate Glubish called the project a big deal for Alberta and said the province had built a regulatory framework to attract data center investment; Alberta is prioritizing hyperscale AI datacenters that build or secure their own power because the provincial grid cannot support multiple large facilities simultaneously. Meta said the datacenter will use a closed-loop cooling system that will not draw water from surrounding sources, and will invest an additional $42 million in local roads and water infrastructure.

Read at AP News ↗ • Read at Bloomberg ↗ • Read at Global News ↗

U.S. rare earths from federally backed producers are flowing to Japan and South Korea, not to U.S. buyers

Rare earth products from MP Materials, Energy Fuels and Phoenix Tailings, all recipients of billions of dollars in U.S. government support, are being sold to Asian customers because domestic magnet manufacturing capacity has not scaled to absorb the output, per Ars Technica. Phoenix Tailings CEO Nick Myers said Japanese customers are demanding the materials, and warned that unless U.S. defense primes move quickly he will sell out to other customers paying top dollar faster. MP Materials, the largest U.S. producer, generated its neodymium-praseodymium sales mainly through Sumitomo Corporation of Americas, which distributes to Japanese customers, per MP's latest quarterly earnings; MP has stopped selling to China's Shenghe Resources under its U.S. government deal and has struck agreements with General Motors and Apple to supply magnets. China retains its lock on global rare earth supplies, which Beijing has restricted access to this year.

Read at Ars Technica ↗

VI.Global & Geopolitics

IMF raises South Korea 2026 growth forecast most among 30 major economies, citing AI hardware exports

The International Monetary Fund raised South Korea's 2026 growth forecast to 2.6%, up 0.7 percentage points from its April projection and the largest upgrade among 30 major economies covered in the July update, per Bloomberg. The IMF attributed the upgrade to stronger than expected semiconductor and AI hardware exports, which have outweighed higher energy costs from the Middle East conflict, per Korea JoongAng Daily and the Dong-A Ilbo. First-quarter South Korean growth reached 7.5% annualized, far above the IMF's April forecast of 1.8%. The IMF identified South Korea, Taiwan, Thailand and Malaysia as the world's four largest net exporters of AI hardware, and raised its 2027 South Korea growth forecast to 2.5% from 2.1%.

Read at Bloomberg ↗ • Read at Korea JoongAng Daily ↗ • Read at The Dong-A Ilbo ↗

Bank of England's Financial Policy Committee flags AI driven equity leverage even as it moves to loosen bank capital rules

The Bank of England's Financial Policy Committee said Tuesday it plans to remove and loosen post-2008 leverage ratio buffers in a way that would primarily benefit the largest UK banks and building societies, including NatWest, Lloyds, Nationwide and Santander UK, per The Guardian. Some FPC members warned the proposal might amplify current risks by driving an unwanted increase in market-based leverage, and pointed to a surge in debt-fueled AI stock purchases as a specific concern. The FPC will complete a review of any financial stability gaps by the end of September, before a broader consultation package goes out in early 2027. Forbes reported the BoE's stance fits a broader pattern of central banks engaging with AI bubble questions, and said the bank warns AI's stability impact could surpass previous tech waves.

Read at The Guardian ↗ • Read at Forbes ↗