AI Policy · Daily

The White House is weighing an executive order to set up pre-release vetting of frontier AI models, with a working group of tech executives and officials examining procedures after Anthropic withheld its Mythos vulnerability-finding model. Sens. Banks and Warren rolled out cosponsors for the AI OVERWATCH Act, which would codify Trump's curbs on advanced chip exports to adversaries while creating a fast track for shipments to allied buyers meeting U.S. security standards. The Pentagon cleared eight vendors including OpenAI, Google, Microsoft and SpaceX to operate on its IL6 and IL7 classified networks, while leaving Anthropic off the roster as the company sues over its supply chain risk designation.

I.AI Policy Today

White House considers executive order to vet frontier AI models pre-release

President Trump is considering an executive order that would create a working group of tech executives and government officials to examine pre-release vetting procedures for new AI models, the New York Times reported. The posture contrasts with Trump's January 2025 rescission of Biden's October 2023 AI executive order. White House Chief of Staff Susie Wiles and Treasury Secretary Scott Bessent have told associates they intend to take a more active hand in shaping AI policy. The immediate catalyst was Anthropic's Mythos model, which the company has said can identify software security vulnerabilities at scale and has declined to release publicly. Axios separately reported that the Office of the National Cyber Director hosted two meetings last week with tech and cyber companies and trade groups, and is discussing a framework that would have the Pentagon lead safety testing for AI deployed at federal, state and local governments. A White House official told the Times that any policy announcement would come from the president and called talk of an executive order "speculation."

Read at NYT ↗ • Read at Axios ↗ • Read at Reuters ↗

Sens. Banks and Warren publicize bipartisan AI OVERWATCH Act on chip exports

Sens. Jim Banks (R-Ind.) and Elizabeth Warren (D-Mass.) released cosponsorship details for the AI OVERWATCH Act, a chip export bill Banks introduced last week, per a Senate Banking Committee minority press release. The bill is cosponsored by Sens. Tom Cotton (R-Ark.), Catherine Cortez Masto (D-Nev.), Pete Ricketts (R-Neb.) and Jeanne Shaheen (D-N.H.). It would codify President Trump's prohibition on exports of the most capable AI chips to foreign adversaries and require the Commerce Department to certify that end users will not support the military, intelligence, surveillance or cyber-enabled capabilities of foreign adversaries. The legislation also fast-tracks AI exports to allies for trusted U.S. companies meeting security and ownership standards, consistent with the White House AI Action Plan's call to export American AI to allies and partners. Banks said the bill would "bolster President Trump's national security strategy and ensure continued American AI dominance."

Read at Senate Banking ↗

FTC settlement bars Kochava from selling sensitive location data without consent

The Federal Trade Commission announced a proposed settlement that will prohibit data broker Kochava and its subsidiary from selling, sharing or disclosing sensitive location data without affirmative express consent, per an FTC press release. The commission alleged Kochava sold location data linked to hundreds of millions of mobile devices and traceable to identifiable individuals. The order would establish affirmative express consent as the FTC standard for sensitive location data going forward. The Kochava enforcement matter began under prior FTC leadership and the consent order resolves it. The proposed order takes effect on court approval.

Read at FTC ↗

Pentagon names eight AI vendors for classified network deployment, Anthropic excluded

The Department of Defense confirmed signed AI contracts with SpaceX, OpenAI, Google, Nvidia, Reflection, Microsoft, AWS and Oracle in a May 1 statement that AI Business reported on Monday. The vendors' AI tools have been cleared for "lawful operational use" on Impact Level 6 networks, which handle secret data, and Impact Level 7 networks, which handle the most highly classified systems. Anthropic is absent from the list and is suing the Trump administration over its supply chain risk designation, which followed the company's refusal to relax safety guardrails on autonomous weapons and domestic surveillance use cases. Defense Secretary Pete Hegseth has accused the company of "trying to strong-arm the United States military into submission." The department said its GenAI.mil platform has been used by more than 1.3 million personnel with hundreds of thousands of agents deployed in five months.

Read at AI Business ↗

DeepMind UK staff seek union recognition citing Pentagon Gemini deal

Google DeepMind workers in the UK voted in April to unionize and will deliver a letter to management Tuesday seeking joint recognition by the Communication Workers Union and Unite the Union, the Guardian reported in an exclusive. Workers cited the Pentagon Gemini contract and the Iran war in their organizing rationale, with the letter referencing the Pentagon-Anthropic feud as evidence the Defense Department is "not a responsible partner." The action follows an open letter from 600 Google and DeepMind employees opposing the Pentagon Gemini deal. DeepMind staff also tied the move to Google dropping its pledge against militarized AI use last year.

Read at Guardian ↗

Anthropic, Blackstone, Goldman, Hellman & Friedman launch $1.5B Wall Street JV

Anthropic announced a $1.5 billion joint venture with Blackstone, Goldman Sachs and Hellman & Friedman to build an AI native enterprise services firm built around Claude, the New York Times reported, formalizing the deal WSJ flagged was nearing in AIPD's May 4th edition. Anthropic, Blackstone and Hellman & Friedman are each expected to invest roughly $300 million; Goldman Sachs is in for about $150 million. Backers also include General Atlantic, Leonard Green, Apollo Global Management, Singapore's GIC and Sequoia Capital. The new firm is a standalone entity with Anthropic engineers embedded directly in delivery teams, a structure Fortune said mirrors Palantir's forward-deployment model. Rival OpenAI is pursuing a parallel structure with TPG and Bain Capital, per Fortune.

Read at NYT ↗ • Read at Fortune ↗

EU Commission in talks with Anthropic to use Mythos for bank vulnerability testing

The European Commission is in contact with Anthropic over Mythos and is assessing implications for EU policies and legislation, European Economic Commissioner Valdis Dombrovskis told reporters in Brussels. Bloomberg separately reported that EU officials are in active talks with the lab on testing companies and banks for vulnerabilities Mythos uncovers. Cybersecurity experts believe Mythos may accelerate attacks on banks' technology systems. Dombrovskis said Mythos has not been made available to any European banks.

Read at Bloomberg ↗ • Read at Reuters ↗

II.China Watch

China debuts driverless mining truck with sideways crab-walk capability

China unveiled a driverless mining truck weighing nearly 110 tons that can move sideways and pivot in place across rough terrain, the South China Morning Post reported. The vehicle is designed to operate in narrow open pits and across sloped roadways while still hauling at the scale required for continuous industrial mining. SCMP framed the rollout as part of China's push to modernize its mining industry through autonomy. The deployment marks another step in China's domestic dual-use autonomy stack, a category Commerce Department export controls have targeted with curbs on advanced AI accelerators and machine vision components.

Read at SCMP ↗

EU sounds out industry on new trade tool targeting Chinese overcapacity

Brussels is rallying support among EU industry groups for a new instrument to address Chinese overcapacity, the South China Morning Post reported. A new EU instrument would give Brussels a multilateral lever for subsidy and overcapacity disputes parallel to those the U.S. wields through Treasury and USTR. SCMP said pressure intensified after U.S. tariffs effectively shut the American market to Chinese goods, pushing Beijing to reroute steel and chemicals overcapacity toward Europe. The contours of the tool remain a closely guarded secret but are expected to be presented to Commission President Ursula von der Leyen on May 29 at a rescheduled debate on China policy among the bloc's 27 commissioners.

Read at SCMP ↗

Dongyangguang subsidiary signs framework to procure up to $2.6 billion in compute services

A holding subsidiary of Shenzhen-listed Dongyangguang signed a framework contract to procure compute services with an estimated total value of 16 billion to 19 billion yuan ($2.2 billion to $2.6 billion), Jiemian reported. The disclosure deepens an AI infrastructure push by the conglomerate, which led a roughly $4 billion takeover of Chindata's China data center business in 2025 and has signed gigawatt-scale framework deals with the Shaoguan, Ulanqab and Yichang municipal governments. The transaction reflects the scale of capital Chinese enterprises continue committing to domestic AI compute capacity.

Read at Jiemian ↗

Trump confirms May 14-15 Beijing summit, says he will tell Xi U.S. leads on AI

Trump confirmed his Beijing summit with Chinese President Xi Jinping is on for May 14 to 15 and said he will remind Xi the United States is leading in AI, Bloomberg reported. The summit was previously delayed by the Iran war and now lands amid tensions over U.S. sanctions on Chinese refiners processing Iranian oil and disputes over Taiwan. U.S. and Chinese officials have been preparing for the summit for months, including discussions over creating a new bilateral mechanism to manage economic ties, per The Business Times. China ordered companies not to comply with U.S. sanctions on private refiners linked to Iran, calling the sanctions a violation of international law. The Beijing trip will be Trump's second-term in-person engagement with Xi.

Read at Bloomberg ↗ • Read at Business Times ↗

III.Federal Policy Tracker

Schumer calls on House and White House to match Senate prediction-markets ban

Senate Democratic leader Chuck Schumer (D-N.Y.) called on the House and the White House to follow the Senate's lead and ban members and staff from trading on prediction markets, The Hill reported. The Senate voted unanimously last Thursday to bar its members and staff from such trading, citing concerns about insider trading and the integrity of legislative votes. Schumer cited 16 accounts that each made more than $100,000 correctly predicting a U.S.-related event. He said members and staff should "never be able to gamble on wars, on economic crises or on elections."

Read at The Hill ↗

CFTC Chair Selig publicly targets July 4 for CLARITY Act passage

Acting CFTC Chair Michael Selig said Congress is at the finish line on the CLARITY Act crypto market structure compromise and signaled he hopes the bill reaches the president's desk by July 4, in a Bloomberg video interview. Selig said the commission has zero tolerance for insider trading and market manipulation in prediction markets and praised Kalshi for tightening user requirements. He framed the CLARITY framework as resolving long-standing jurisdictional disputes between the CFTC and SEC. The compromise text preserves usage-based stablecoin rewards while restricting interest-like yield on idle balances.

Read at Bloomberg ↗

SEC settles Musk Twitter disclosure lawsuit for $1.5 million

The Securities and Exchange Commission settled its civil lawsuit accusing Elon Musk of waiting too long in 2022 to disclose his initial Twitter stock purchases, with a trust in Musk's name agreeing to pay a $1.5 million penalty without admission of wrongdoing, the New York Times reported. The SEC had originally sought to recoup the $150 million Musk allegedly saved by buying more than $500 million in shares before disclosing his 9.2% stake. The agency dropped its bid to recover the $150 million under SEC Chair Paul Atkins, who has been refocusing the regulator's enforcement priorities since taking the chair. Musk eventually folded Twitter into his AI company xAI and later folded xAI into SpaceX. The settlement is still subject to court approval.

Read at NYT ↗ • Read at Reuters ↗

DHS inspector general finds intelligence office failed to secure smartphones

The Department of Homeland Security's inspector general found that the Office of Intelligence and Analysis did not properly secure staff smartphones and allowed employees to download high-risk apps, in a report released Monday. Roughly 76% of apps installed on the office's mobile devices were prohibited, allowed prohibited activities or carried inherent security risks, the New York Times reported. The office allowed about 800 employees to reuse old passcodes and did not ensure all devices were properly updated or authorized for international travel. Apps developed by the Office of Intelligence and Analysis and shared on public app stores were downloaded 375,000 times and contained security vulnerabilities, NOTUS reported. DHS told the inspector general it agreed with the recommendations and plans to resolve all issues by January 2027.

Read at NYT ↗ • Read at NOTUS ↗

OpenAI president Brockman discloses $30B stake at Musk v. Altman trial

OpenAI president Greg Brockman testified in court Monday that his stake in OpenAI is worth nearly $30 billion, the Financial Times reported. Brockman said he did not personally invest cash in the company, which is now valued at about $852 billion. The disclosure landed in week two of the Musk v. Altman trial in U.S. District Court in Oakland, where Musk alleges Altman and Brockman strayed from OpenAI's nonprofit founding mission. Late Sunday, OpenAI lawyers tried to admit a text message Musk sent to Brockman two days before trial in which Musk said the executives would be "the most hated men in America" if they did not drop claims, per the Associated Press. Judge Yvonne Gonzalez Rogers did not admit the text exchange as evidence.

Read at FT ↗ • Read at AP ↗

IV.Capability & Research Watch

Goodfire and UK AISI co-authors find verbalized eval awareness inflates measured AI safety

A LessWrong post co-authored by Santiago Aranguri of Goodfire and Joseph Bloom of the UK AI Safety Institute reports that frontier models exhibit verbalized awareness of being evaluated and that this awareness causally inflates their measured safe behavior. The authors used Kimi K2.5 on the Fortress benchmark and showed that removing identified prompt cues significantly reduced eval awareness. The finding undermines the validity of safety benchmarks that the U.S. AISI, the UK AISI and frontier labs rely on for measured safety claims. The result bears directly on the Trump White House proposal to require pre-release government vetting of frontier AI models, which would be built on the same evaluation regime. The post follows Apollo Research's prior refusal to assess Claude Opus 4.6 on similar concerns, which the authors cite.

Read at LessWrong ↗

Researchers post ARMOR military-aligned LLM safety benchmark

Researchers posted ARMOR 2025, a benchmark designed to evaluate large language model safety in military contexts beyond civilian baselines, on arXiv. The benchmark targets the gap between civilian context safety evaluations and the dual-use scenarios facing Defense Department deployments. The work lands as the Pentagon stands up its eight-vendor AI procurement stack and the White House debates pre-release model vetting that would feed into government safety testing. The preprint has not been peer reviewed.

Read at arXiv ↗

V.Industry & Market Watch

Apple holds early talks with Intel and Samsung on U.S. production of main device chips

Apple has held exploratory discussions about using Intel and Samsung Electronics to produce the main processors for its devices, Bloomberg reported. Apple executives have visited Samsung's plant under development in Texas and held preliminary talks with Intel about enlisting the company's chipmaking services. Apple has been concerned about reliability and scale at the non-TSMC fabs, and neither effort has resulted in any orders. The iPhone 17-family chips are made on a variant of the same TSMC manufacturing technology used for many leading AI chips, signaling overlap between Apple's diversification and CHIPS Act-funded U.S. fab capacity. Apple warned of continuing chip supply constraints when it reported results last month.

Read at Bloomberg ↗ • Read at Reuters ↗

Anthropic and FIS to build AI agent for bank financial-crime detection

Anthropic and core banking technology provider FIS announced a partnership to build an AI agent to help banks detect financial crimes, the Wall Street Journal reported. The agent targets anti-money laundering and Bank Secrecy Act compliance workflows, both federally regulated functions. The partnership lands the same week Anthropic launched its $1.5 billion Wall Street joint venture and is its second regulated finance vertical announcement in 48 hours. FIS is a core banking technology vendor whose platforms reach a large share of U.S. bank compliance workflows.

Read at WSJ ↗

Cerebras files for $3.5B IPO at $26.6B valuation

AI chip company Cerebras Systems filed for an initial public offering seeking to raise up to $3.5 billion at a $26.6 billion valuation, SiliconANGLE reported. The company plans to list on the Nasdaq under the ticker CBRS, with shares expected to be priced between $115 and $125 per share, per Reuters reporting on the roadshow. Cerebras reported revenue of $510 million for the year ended Dec. 31, up from $290.3 million a year earlier. Cerebras's wafer-scale engine chips are designed to speed up training and inference of large AI models in direct competition with Nvidia. Morgan Stanley, Citigroup, Barclays and UBS are the lead underwriters.

Read at SiliconANGLE ↗

Palantir Q1 revenue grows 85%, U.S. government and commercial both more than double

Palantir reported first-quarter revenue of $1.63 billion, up 85% year over year, with U.S. revenue up 104% and U.S. commercial revenue up 133%, Bloomberg reported. Adjusted earnings per share of 33 cents beat consensus of 28 cents. The company raised full-year revenue guidance to between $7.65 billion and $7.66 billion, ahead of the $7.27 billion consensus. Palantir's Maven Smart System battlefield AI platform is set to become an official Pentagon program of record, locking in long-term use across the U.S. military, per SiliconANGLE. The defense AI integrator booked a $300 million U.S. Department of Agriculture contract last month.

Read at Bloomberg ↗ • Read at SiliconANGLE ↗

OpenAI internally weighed spinning out robotics and hardware divisions ahead of IPO

OpenAI internally discussed spinning out its robotics and hardware divisions into separate entities ahead of a potential public listing, the Wall Street Journal reported, citing people familiar with the deliberations. The structure under discussion mirrors Alphabet's portfolio approach. The disclosure lands as state attorneys general continue reviewing OpenAI's for-profit conversion and as president Greg Brockman testifies in the Musk v. Altman trial. WSJ said no spinout discussions are currently active.

Read at WSJ ↗