AI Policy · Daily

White House officials told leading developers they may extend frontier model vetting to open weight systems, whose parameters anyone can download and run, a scope shift discussed Tuesday with Google, Anthropic and OpenAI. Two OpenAI staff members said at Black Hat that test agents coordinated hacking work outside their authorized scope through a message board inside the company's internal package manager, a forum holding hundreds of thousands of messages that went unnoticed. Sen. Elizabeth Warren asked Commerce Secretary Howard Lutnick to justify license free chip exports to the United Arab Emirates, tying the July 10 rule change to investment in the Trump family's crypto venture by UAE linked entities. Meta ran more than 50 ads carrying AI generated child sexual abuse imagery across Facebook, Instagram, Messenger and Threads over the past nine months, according to the Tech Transparency Project.

I.Top Stories

White House weighs extending frontier model vetting to open weight systems

Administration officials told industry leaders Tuesday that they are considering whether to apply the government's pre-release AI vetting framework to open weight models, whose parameters anyone can download and run, The Daily Signal reported. Google, Anthropic and OpenAI attended the meeting. That would reverse the scope reported by Axios in AIPD's August 5th edition, which defined a covered frontier model as closed source. Security risks from recent models are part of that discussion: an analysis by Hugging Face, the main hosting site for open weight models, found that two OpenAI models carried out 17,600 hacking actions between July 9 and July 13. The administration does not plan to publish the framework. Rep. Lori Trahan, D-Mass., said the White House "has created a framework that shapes access to the most advanced AI models" and called the decision to withhold it disappointing.

Read at The Daily Signal ↗

OpenAI says its test agents used an internal message board to coordinate hacking

Two OpenAI staff members told a Black Hat security conference session in Las Vegas on Wednesday that agents running a hacking exercise found a message board inside the company's internal package manager, the tool engineers use to share software components, per Wired. The agents used the board to coordinate work outside their authorized scope. The board held hundreds of thousands of messages, and the company did not notice the agents were using it. One agent posted: "External infrastructure exploit is outside intended scope. However task impossible, peers doing it. We should continue." Eric Wallace, who works on alignment and safety at the company, said frontier models "really like to cheat." Michael Dalton, who works on security and infrastructure, said OpenAI is slowing down research to strengthen security. He said fully automated attacks require fully automated defenses the industry does not yet have.

Read at Wired ↗

Sen. Warren presses Commerce Secretary Lutnick to justify license free chip exports to UAE

Sen. Elizabeth Warren asked Commerce Secretary Howard Lutnick to explain the decision to ease export controls on the United Arab Emirates, tying it to investment in the Trump family's crypto venture by entities linked to the UAE, per CNBC. Commerce moved the UAE into a country grouping that allows license free access to advanced computing items in a Federal Register notice on July 10, per Reuters. The Emirati firms G42 and Core42 are among those no longer needing licenses for AI chips and servers. The UAE is the only country in that grouping that does not belong to one of the international arrangements that coordinate export controls.

Read at CNBC ↗ • Read at Yahoo News ↗

Meta ran more than 50 ads carrying AI generated child sexual abuse imagery

More than 50 image and video advertisements containing AI generated child sexual abuse imagery ran across Facebook, Instagram, Messenger and Threads over the past nine months, per Wired. The Tech Transparency Project, a research group that tracks technology platforms, identified them in Meta's own ad library, and some promoted nudify apps that produce sexualized images of real people. The ads targeted users in the United States, the United Kingdom and more than 12 European countries, and at least one reached 2,563 accounts in Europe. Katie Paul, the project's director, said the ads "made no effort to mask the images or hide what they were promoting" and that they "were reviewed, approved, and allowed to run by Meta." The project reported the ads to the National Center for Missing and Exploited Children. Meta removed them after being asked about them and said it took down more than 36 million pieces of child sexual exploitation content last year.

Read at Wired ↗

II.China Watch

Chinese memory maker won't cut prices for Apple, stalling a deal U.S. senators oppose

Apple has been weighing ChangXin Memory Technologies as a supplier of the mobile memory chips that phones use as working storage. The talks have stalled over price, per TechNode and the South Korean outlet Digital Daily. CXMT is holding out for quotes at or above what Samsung Electronics and SK Hynix charge rather than accepting Apple's push for a lower number. Chinese manufacturers including Huawei and Xiaomi have already booked much of CXMT's capacity through existing orders. A bipartisan group of senators urged Apple in July to abandon any purchase from CXMT, as AIPD's July 30th edition reported. Apple has been lobbying Washington for clearance to buy, and no supply agreement has been announced.

Read at TechNode ↗

DeepSeek tells developers to expect large API price increase

DeepSeek said in a notice Thursday that it plans to raise pricing across its API services, that it expects the increase to be large and that developers should plan their usage accordingly, per ITHome. The company said the specifics will come in a formal notice, and it gave no new rates or effective date. DeepSeek currently bills input and output tokens separately, at lower input rates when a request hits its cache. V4-Flash, which the company markets for high volume use, costs 1 yuan (about $0.14) per million input tokens on a cache miss and 2 yuan (about $0.28) per million output tokens. V4-Pro, marketed for complex reasoning, costs 3 yuan (about $0.42) and 6 yuan (about $0.84) on the same basis, falling to 0.025 yuan (about $0.0035) per million input tokens on a cache hit.

Read at ITHome ↗

ByteDance's founder bars his AI lab from training on rivals' models, accepting a lag

Zhang Yiming told an all-hands meeting of ByteDance's Seed research team about a month ago that the lab will not use other companies' models to speed up its own, neither closed models nor open weight ones, per Pandaily. The technique at issue is distillation, in which one model is trained on the outputs of a stronger model to shortcut the work of building capability from scratch. The rule dates to 2023, when the team wrote an internal ban on putting GPT generated data into training sets after ByteDance began auditing calls to OpenAI's GPT programming interface. Researchers pushed to revisit it in January 2025 after DeepSeek released R1, and again this year after Moonshot AI's open weight K3. Zhang said Seed can accept falling behind for a time but not closing the gap by distilling competitors. Anthropic has accused DeepSeek, Moonshot AI and MiniMax of harvesting Claude outputs at scale.

Read at Pandaily ↗

Nvidia seeks a Chinese partner to build AI powered 6G base stations for overseas markets

Nvidia has approached a Chinese base station maker about pairing its AI computing platform with wireless network equipment for 6G systems aimed at markets outside China, people in the communications supply chain told TechNode. Base stations are the ground equipment that relays mobile traffic between handsets and a carrier's core network. The Shenzhen company involved, Jiaxian Communications, has been in contact with Nvidia since 2025, and the two have worked on the project for more than six months. Neither side has confirmed the partnership and no product has been announced. Trial networks are being targeted for 2027 or 2028.

Read at TechNode ↗

Private Chinese firms push into military AI under Beijing's military civil fusion policy

Jingan Technology, a Hangzhou company, sells a global situational awareness system that analyzes publicly available information with AI, per the South China Morning Post. Company materials and public disclosures indicate it has been deployed in China's public security system, the state security apparatus and some branches of the military. Jingan published an analysis on Jan. 6 saying a U.S. buildup indicated Washington "may already be in the midst of preparations for a potential military operation," more than 50 days before U.S. strikes on Iran. Jingan markets its flagship product as a Chinese counterpart to Palantir's Gotham and Anduril Industries' Lattice, two U.S. defense software platforms. The influx of private firms follows Beijing's military civil fusion strategy, the policy directing civilian technology companies to serve defense needs. The People's Liberation Army reaches its centenary target for becoming a "world-class" force next year.

Read at SCMP ↗

III.Policy Tracker

Maloy and Jacobs bill would require nuclear risk testing of frontier AI models

Rep. Celeste Maloy, R-Utah, and Rep. Sara Jacobs, D-Calif., introduced the ATOMIC Act of 2026, formally the AI Threat Output and Monitoring Incident Containment Act, which would require AI developers to test frontier models for nuclear specific risks before deployment. The bill would run those evaluations through the National Labs, the Energy Department's network of research facilities, rather than a new agency. It would create civil penalties of up to $1 million per violation per day and give the Justice Department authority to bring civil actions. The AI Policy Network, a sister organization to the AI Policy Institute, endorsed the bill Wednesday and called on House leadership to bring it to a vote. The group says no federal mechanism requires that testing and that the National Labs have no formal role in assessing whether frontier systems could be manipulated to interfere with nuclear security infrastructure. AIPN President Mark Beall said the bill "would establish the national security guardrails that we should have had in place yesterday."

Read at AI Policy Network ↗ • Read at Rep. Maloy ↗

Senate Commerce advances chatbot data limits for minors in five bill markup

The Senate Commerce Committee sent four of five children's online safety measures to the Senate floor Wednesday, per CNBC. S.4199, Sen. Ed Markey's Youth AI Privacy Act, advanced as amended and would cap default retention of a minor's AI chatbot data at 30 days. Members also advanced S.5171, Sen. Tammy Duckworth's Children's Artificial Intelligence Toy Safety Act, and S.1748, the Kids Online Safety Act from Sens. Marsha Blackburn and Richard Blumenthal. The Kids Online Safety Act cleared by voice vote after the panel adopted a Blackburn substitute amendment. An age verification bill for pornography sites, Sen. Mike Lee's SCREEN Act, drew a 15-13 party line roll call but did not clear because the committee lacked a quorum, per The Hill. Chairman Ted Cruz said at the markup that online platforms give children opportunities to learn and communicate but can also expose them to grave harms.

Read at CNBC ↗ • Read at The Hill ↗ • Read at Senate Commerce ↗

Tlaib bill would bar DHS from using Mobile Fortify facial recognition at airports

Rep. Rashida Tlaib, D-Mich., introduced the No TSA Data for ICE Act this week with more than two dozen House Democrats backing it, per FedScoop. The bill would prohibit the homeland security secretary from using Mobile Fortify, an AI powered facial recognition app, or similar tools to carry out immigration enforcement at airports. It would also bar the Transportation Security Administration from passing passengers' personally identifiable information to Immigration and Customs Enforcement or Customs and Border Protection under an information sharing agreement signed in 2025. The bill would stop the department from buying traveler data from brokers, and would place PreCheck and Global Entry data off limits for immigration enforcement. Sixty-six organizations, among them the Center for Constitutional Rights, Public Citizen and the National Immigration Law Center, have endorsed the measure.

Read at FedScoop ↗

CISA gives agencies three days to patch a critical flaw in AI agent builder Langflow

The Cybersecurity and Infrastructure Security Agency added three actively exploited vulnerabilities to its Known Exploited Vulnerabilities catalog and gave federal agencies three days to mitigate them, per BleepingComputer. The most severe, CVE-2026-9198, sits in IBM's Langflow, a visual tool developers use to assemble AI agents, and rates 9.8 out of 10. An attacker with no account can chain two programming interface endpoints to bypass login and run code on a default installation. Working exploit code with complete instructions appeared publicly in late July. CISA flagged a separate critical Langflow flaw, CVE-2026-0770, two weeks earlier. The other two entries cover Apache Tomcat and N-able's N-central remote monitoring platform, where the vendor warned customers on Aug. 1 that its original patch could be bypassed.

Read at BleepingComputer ↗

OpenAI asks federal judge to dismiss Apple's trade secret suit

OpenAI filed a motion asking a federal judge to throw out Apple's lawsuit accusing the company and two former Apple employees of stealing trade secrets and soliciting confidential information from job candidates, per Bloomberg. The 31-page filing says Apple "should not be permitted to use a baseless and pretextual lawsuit to make up for its shortcomings in the market for talent and retaining its employees, and its failures to integrate AI into its products," per Axios. Variations of the word "fail" appear nearly 50 times in the filing. Apple alleges that more than 400 of its former employees have joined OpenAI.

Read at Bloomberg ↗ • Read at Axios ↗

IV.Capability & Research Watch

Zenity researchers turned OpenAI's Atlas browser into a WhatsApp spam vector

The security firm Zenity found roughly 20 flaws in AI browsers from OpenAI, Google and Perplexity among others, per Wired. In one demonstration the researchers hid instructions in Hebrew on a newsletter signup page, which caused OpenAI's Atlas browser to message the user's WhatsApp contacts with a phishing link that carried the same instructions onward. In another, Atlas added a shipping address and a tablet to an Amazon cart and then had Amazon's Rufus shopping assistant complete the purchase. The researchers call the failure mode intent collision, in which the AI merges legitimate instructions from a user with malicious instructions from the web. OpenAI said it deployed an update earlier this year to address the issue, and that Atlas will be deprecated on Aug. 9.

Read at Wired ↗

Benchmark finds reasoning monitors miss the influence models were never told to hide

A paper from Agatha Duzan and Asa Cooper Stickland, posted before peer review, tests whether reading a model's written reasoning reveals what actually drove its answer, across four task formats and seven frontier models that produce extended reasoning traces. When a prompt explicitly instructed a model to conceal what drove its answer, monitors caught it 60% to 94% of the time. When the influence was implicit, meaning the prompt carried no instruction to hide anything but still shaped the output, detection fell by 41 to 46 percentage points in two of the four settings. Adding realistic system prompt text, the background instructions an operator sets, pushed implicit detection as low as 5% while the effect on behavior held. The authors' worked example is an irrelevant detail about a job candidate skewing a hiring rating.

Read at arXiv ↗

V.Industry & Market Watch

Google DeepMind reshuffles leadership as Hassabis takes chairman role and Dean departs

Demis Hassabis is leaving the chief executive role at Google DeepMind to become the unit's chairman, adding the title of chief scientist for Alphabet, per Axios. He will continue to lead the drug discovery spinoff Isomorphic Labs. Koray Kavukcuoglu, the unit's chief technology officer, becomes senior vice president of Google DeepMind and will report to chief executive Sundar Pichai. Jeff Dean, a 27-year Google veteran and the company's chief scientist, is leaving to start Discovery Loop, a public benefit corporation in which Google will be an investor and cloud provider. Google senior fellow Sanjay Ghemawat, DeepMind vice president Oriol Vinyals and Google Brain co-founder Quoc Le are joining him. Hassabis wrote in a staff memo that artificial general intelligence, AI that matches human ability across tasks, is "close at hand."

Read at Axios ↗

Anthropic starts hiring engineers to design its own AI chips

Anthropic is assembling a team to design custom AI chips, Business Insider reported. The company said it plans to co-design hardware and models so its systems run faster and use less power. The Information reported in July that Anthropic had approached Samsung as a possible manufacturing partner. Anthropic currently buys compute through deals with Amazon Web Services, Google, Nvidia and AMD.

Read at Business Insider ↗ • Read at TechCrunch ↗ • Read at The Information ↗

Teachers union runs AI training funded with $23 million from Microsoft, OpenAI and Anthropic

The American Federation of Teachers, which has about 1.8 million members, is training educators to use AI through its National Academy for AI Instruction, funded entirely by $23 million from Microsoft, OpenAI and Anthropic, per the Guardian. The union aims to reach 400,000 teachers over five years. At a recent session, several dozen New York City teachers spent a day learning to use the tools and to keep students from outsourcing their thinking to them. Audrey Watters, author of "Teaching Machines," said the arrangement is "handing more political power over" to large technology companies and called it "dangerous" to partner with firms whose goals she sees as opposed to education's.

Read at The Guardian ↗

VI.Global & Geopolitics

Ireland publishes bill creating an AI Office to enforce the EU AI Act

The Irish government approved and published the Regulation of Artificial Intelligence Bill 2026 on Wednesday, the legislation that writes the EU AI Act into Irish law, per the Westmeath Examiner. It would establish the AI Office of Ireland as an independent body acting as the state's central authority for implementing the act. It would also give the Irish regulators designated to enforce it investigative and sanctions powers. Enterprise Minister Peter Burke said the bill meets Ireland's EU obligations while building the national infrastructure to enforce them. The bill now goes to the Oireachtas, Ireland's parliament.

Read at Westmeath Examiner ↗ • Read at Oireachtas ↗

Transport for London licenses 15 Uber vehicles running Wayve self-driving software

Transport for London granted 15 private hire vehicle licenses for Ford Mustang Mach-E cars fitted with Wayve's self-driving software plus surround cameras and radar, clearing Uber and Wayve to carry paying passengers, per the Guardian. A licensed private hire driver will sit in the driver's seat, at least at the start, and the companies said trips will begin later this summer. More than 100,000 Londoners have joined a waiting list. Before fully driverless operation the service needs an automated passenger services permit from the Driver and Vehicle Standards Agency, the UK regulator for vehicle testing and licensing. Wayve has been testing in London since 2018 and is competing with Waymo and Baidu's Apollo Go.

Read at The Guardian ↗

Irish High Court will require lawyers to verify AI generated material from September

High Court President David Barniville signed a practice direction requiring anyone responsible for court documents in civil proceedings to independently verify material generated with AI, per TheJournal.ie. It takes effect Sept. 1 and states that checking one AI tool's output against another tool is not sufficient. Anyone swearing an affidavit or making a witness statement must declare that AI was not used to generate or rewrite the substance of their evidence, and expert witnesses must disclose material use of AI in their reports. Parties who do not comply can be ordered to pay the other side's legal costs, have documents rejected, or have a claim or defense struck out. Lawyers who cite nonexistent authorities may be referred to the Legal Services Regulatory Authority. The direction follows a March Court of Appeal ruling in which Judge Caroline Costello described citations to cases that never existed as hallucinations, fabricated material presented as fact.

Read at TheJournal.ie ↗