House advances China AI curbs | GOP privacy preemption bills | CISA nominee out

Bottom Line Up Front

The House Foreign Affairs Committee advanced bipartisan bills extending export controls beyond chips to cover AI models and technology transfers to China, pushing past the Trump administration's reluctance to expand curbs and building on earlier GOP efforts to sanction Chinese model distillation.
The Pentagon's FY27 request earmarks $54 billion for the Defense Autonomous Warfare Group, a roughly 24,000% jump that funds autonomous drones, AI integration, and Golden Dome, with Senate Armed Services Strategic Forces set to scrutinize missile defense on April 27.
Reps. Guthrie and Joyce introduced the SECURE Data Act and GUARD Financial Data Act, paired bills creating a federal privacy standard that would preempt California, Colorado, Connecticut, and Virginia laws while omitting a private right of action critics say weakens consumer enforcement.
CISA nominee Sean Plankey withdrew after 13 months without a Senate vote, extending the civilian cyber agency's leadership vacuum; separately, reporting surfaced that Anthropic's Mythos deployment across federal agencies has bypassed CISA entirely.

I.AI Policy Today

House Foreign Affairs Committee advances bipartisan AI export-control bills targeting China

The House Foreign Affairs Committee advanced a slate of bipartisan export-control bills Wednesday aimed at restricting AI technology transfers to China, Bloomberg reported. The measures extend the export-control framework beyond the semiconductor restrictions currently administered by the Bureau of Industry and Security to cover AI models and technology. They also include proposals to raise civil penalties for export-control violations and create a whistleblower incentive program. The committee action builds on House Republicans' effort to sanction Chinese AI model distillation. The Trump administration has held back on imposing new curbs beyond existing chip controls.

Read at Bloomberg ↗

Pentagon FY27 request includes $54 billion for Defense Autonomous Warfare Group

The Pentagon's fiscal year 2027 budget request includes roughly $54 billion for the Defense Autonomous Warfare Group (DAWG), a roughly 24,000% increase over DAWG's approximately $225.9 million FY26 allocation, the Guardian reported. The allocation sits within the $1.5 trillion defense budget the Trump administration submitted to Congress this week. Budget documents fund AI integration, autonomous drones, and the Golden Dome missile defense system. DAWG will replace the Replicator program as the Pentagon's focal organization for drone and autonomy capabilities. The Senate Armed Services Subcommittee on Strategic Forces will examine missile defense activities at a hearing scheduled for April 27.

Read at The Guardian ↗ • Read at Congress.gov ↗

House Republicans introduce paired bills to preempt state data privacy laws

Energy and Commerce Committee Chair Brett Guthrie (R-Ky.) and Rep. John Joyce (R-Pa.) introduced the SECURE Data Act and the GUARD Financial Data Act on Wednesday, CNBC reported. The paired bills would create a federal data privacy standard that preempts state privacy laws, including those in California, Colorado, Connecticut, and Virginia. Critics say the bills omit a private right of action, making the federal standard weaker than existing state regimes on consumer enforcement. The SECURE Data Act moves through House Energy and Commerce, where Guthrie serves as chair, while the GUARD Financial Data Act moves through House Financial Services.

Read at CNBC ↗ • Read at The Hill ↗

CISA director nominee Sean Plankey withdraws after 13 months without a confirmation vote

Sean Plankey sent a letter to President Trump on Wednesday asking him to withdraw his nomination to lead the Cybersecurity and Infrastructure Security Agency, CyberScoop reported. Plankey had waited 13 months without receiving a Senate confirmation vote. The withdrawal extends a leadership vacancy at the federal government's lead civilian cybersecurity agency, where Nick Andersen is serving as acting director following Madhu Gottumukkala's departure. No replacement nominee has been announced.

Read at CyberScoop ↗

Anthropic's Mythos rollout has excluded CISA from access

Several U.S. federal agencies are using Anthropic's Mythos model for vulnerability assessment, but the Cybersecurity and Infrastructure Security Agency is not among them, the Verge reported. The NSA and Commerce Department are among the agencies testing Mythos; national cyber director Sean Cairncross is negotiating broader civilian access. The gap leaves the federal government's lead civilian cybersecurity coordinator without access to a frontier AI tool that other agencies and private-sector institutions are deploying. The same day, Plankey withdrew his CISA director nomination.

Read at The Verge ↗

USDA awards Palantir $300 million blanket purchase agreement for IT modernization

The Department of Agriculture awarded Palantir Technologies a $300 million blanket purchase agreement covering IT modernization and digital services, FedScoop reported. The deal continues the USDA's "One Farmer, One File" data initiative, under which Palantir has consolidated farmer data into a single analytics platform. The BPA structure allows the USDA to issue rolling task orders under the agreement. Agriculture Secretary Brooke Rollins previewed the initiative at a February 2026 event.

Read at FedScoop ↗ • Read at CNBC ↗

Sen. Hawley warns Republicans face "political cost" for backing AI industry

Sen. Josh Hawley (R-Mo.) warned fellow Republicans that the party faces a "political cost" for failing to rein in Big Tech and AI, the Financial Times reported. Hawley described the AI industry's influence campaign on Capitol Hill as a $300 million lobbying effort. He did not name specific companies or legislation in the interview. A separate advocacy group analysis found 11 major tech firms spent a combined $226,000 per day on congressional lobbying in Q1 2026, nearly double the rate in 2020, according to Fortune.

Read at FT ↗ • Read at Fortune ↗

II.China Watch

MIIT reports China's domestic computing power has reached 1,882 exaflops, far above Top500 tallies

China's Ministry of Industry and Information Technology reported the country's computing power has reached 1,882 exaflops, more than 6,000 times higher than China's figure in the Top500 global supercomputing ranking, SCMP reported. SCMP noted the two metrics are not directly comparable, describing the gap as China's "dark compute power." The MIIT total rolls up distributed and non-centralized compute that does not appear in standard data-center tallies.

Read at SCMP ↗

Pudu Robotics opens Dallas U.S. headquarters, closes ~$150M round at $1.5B valuation

Chinese service-robotics firm Pudu Robotics inaugurated its new U.S. headquarters in Richardson, Texas on April 22, relocating from Santa Clara, per Pandaily and Dallas Innovates. The company announced a funding round of nearly $150 million, pushing its valuation above $1.5 billion, with proceeds earmarked for embodied-AI development and scaled manufacturing, per its PR Newswire release. A China domiciled, embodied AI vendor expanding its commercial footprint on U.S. soil tees up CFIUS and Entity List questions parallel to those already raised over Chinese AI-chip firms. The new Richardson site combines office space, product showroom, and on-site warehousing to anchor Americas sales and operations.

Read at Pandaily ↗

Ant Group releases Ling-2.6-Flash, a 104B-parameter MoE model live on OpenRouter

Ant Group's Inclusion AI team launched Ling-2.6-Flash on April 22, an instruct-tuned Mixture-of-Experts model with 104 billion total parameters and 7.4 billion active, optimized for token efficiency, per Pandaily. The model is available through Alipay's Tbox platform and went live on OpenRouter, having previously circulated under the codename "Elephant Alpha." Distribution through OpenRouter places a Chinese frontier-tier MoE directly on a U.S.-headquartered inference gateway, within scope of the House Select Committee on the CCP's recent inquiries into model-hosting supply chains. Ant Group says the model leads models of its size on agent benchmarks including BFCL-V4, TAU2-bench, and SWE-bench Verified.

Read at Pandaily ↗

Anker debuts in-house "Thus" compute-in-memory chip for on-device AI

Shenzhen-based consumer-electronics firm Anker unveiled Thus, which it calls the first compute-in-memory AI audio chip with neural-network support, on April 22, per Pandaily and Engadget. The chip embeds computation directly into NOR flash memory cells rather than separating storage and processing, and Anker says it delivers 150 times more AI computing power for noise cancellation than previous soundcore models. The AI silicon architecture sits outside the GPU and NPU categories targeted by U.S. export controls, a chip class that current BIS rules do not expressly address. Anker will debut Thus in a new headphone model at its May 21 "Anker Day" event before rolling it into mobile accessories and IoT devices.

Read at Pandaily ↗

III.Federal Policy Tracker

Sen. Warren seeks details on Palmer Luckey-linked Erebor bank charter approval

Sen. Elizabeth Warren (D-Mass.) formally requested details about the bank approval process for Erebor, a venture linked to Anduril co-founder Palmer Luckey and Palantir co-founder Joe Lonsdale, Bloomberg reported. Warren's letter targets the regulators overseeing the charter process and asks how a defense-tech-adjacent venture secured rapid banking approval. Luckey co-founded Anduril Industries, which holds major Department of Defense AI and autonomous-systems contracts.

Read at Bloomberg ↗

Sen. Blackburn highlights broad support for the TRUMP AMERICA AI Act, creation of a federal AI evaluation program

Sen. Marsha Blackburn (R-Tenn.) highlighted broad support for the TRUMP AMERICA AI Act, which would create an Advanced Artificial Intelligence Evaluation Program within the federal government and permanently authorize the Center for AI Standards and Innovation, per the AI Policy Network. The bill targets superintelligence testing, weaponization risks, loss-of-control scenarios, and cyber threats from frontier AI systems, among other priorities. Advocacy groups supporting safety guardrails, such as the AI Policy Network, commended the bill's national security provisions in a press release.

Read at AIPN ↗

FDA issues first warning letter citing non-compliant AI use in manufacturing

The FDA issued what pharma-compliance analysts describe as the agency's first warning letter targeting non-compliant AI use in current Good Manufacturing Practice (cGMP) production, per ProPharma Group. The enforcement letter addresses governance, validation, and monitoring requirements for AI systems deployed in manufacturing, including the use of AI to generate specifications and master production records without independent quality-unit review. A separate Clarkston Consulting analysis corroborated the FDA action and its emphasis on manufacturing-process validation.

Read at ProPharma ↗

Treasury Secretary Bessent says department canceled Booz Allen Hamilton contracts over IRS leaker vetting

Treasury Secretary Scott Bessent told senators Wednesday that the Treasury Department canceled contracts with Booz Allen Hamilton over the contractor's failure to properly vet the individual who leaked IRS taxpayer data, FedScoop reported. Bessent disclosed the cancellation during testimony to the Senate Appropriations Financial Services and General Government Subcommittee, citing Booz Allen contractor Charles Littlejohn's 2018–2020 leak. Booz Allen Hamilton is one of the federal government's largest AI and data-analytics contractors, with engagements across defense, intelligence, and civilian agencies. The affected Treasury work represented roughly $21 million in obligations and $4.8 million in annual spending.

Read at FedScoop ↗

IV.Capability & Research Watch

North Korean hackers use AI tools to steal up to $12 million in three months

A North Korean state-sponsored group used AI tools for malware development and fake corporate identity creation, stealing as much as $12 million in a three-month campaign, Wired reported. Cybersecurity firm Expel attributed the activity to a group it calls HexagonalRodent and said the operators used AI services from OpenAI, Cursor, and Anima to "vibe code" malware and build phishing infrastructure. Credential-stealing malware was installed on more than 2,000 computers, concentrated among developers working on small cryptocurrency launches, NFT projects, and Web3 work. The AI-assisted techniques enabled operators with limited technical skill to run sophisticated social-engineering and malware campaigns against financial targets.

Read at Wired ↗

Anthropic Claude Code source leak opens AI-era copyright test case

The source code for Anthropic's Claude Code was accidentally exposed earlier this month, creating what the New York Times described as a test case for AI-era copyright law. A University of British Columbia student used AI assistants to rewrite the leaked code in another programming language and repost it, raising new questions about whether machine-assisted rewrites can escape infringement claims. Anthropic has sent DMCA notices to some mirrors, but the code remains available across hundreds of public repositories. The case adds to pending AI-copyright disputes, including litigation from the New York Times and authors' groups against OpenAI.

Read at NYT ↗

V.Industry & Market Watch

Sen. Warren warns AI failure could trigger the next financial crisis

Sen. Elizabeth Warren (D-Mass.) warned at a Vanderbilt Policy Accelerator event that AI-driven capital concentration and market valuations show "striking" parallels to pre-2008 conditions, the Verge reported. Warren said, "I know a bubble when I see one," citing the concentration of capital expenditure among a handful of AI companies and elevated market valuations as specific vulnerabilities. A companion Vanderbilt Policy Accelerator report argued the AI investment cycle is, by one macroeconomic measure, 17 times the size of the dot-com bubble and four times the size of the 2008 housing bubble.

Read at The Verge ↗

Google unveils dedicated TPU chips for AI training and inference

Google announced two new eighth-generation TPUs, the TPU 8t for training and the TPU 8i for inference, in what CNBC described as the company's latest challenge to Nvidia's dominance in AI silicon. The TPU 8i pairs 288 GB of high-bandwidth memory with 384 MB of on-chip SRAM (triple the prior generation) to keep a model's active working set on-chip. The TPU 8t delivers 2.8 times the performance of the seventh-generation Ironwood at the same price, per Google.

Read at CNBC ↗

VI.Global & Geopolitics

London's Metropolitan Police in talks to purchase Palantir AI for criminal investigations

London's Metropolitan Police Service is in procurement talks with Palantir Technologies to acquire AI technology for automating intelligence analysis in criminal investigations, the Guardian reported. The deal would make the Met the largest UK police force to adopt Palantir's analytics platform. Internal Met documents flagged concerns over data handling by a U.S. firm with ties to Immigration and Customs Enforcement and the Israeli military.

Read at The Guardian ↗

Commercial child sexual abuse websites doubled as AI-generated content drives expansion, IWF finds

The Internet Watch Foundation found that the number of commercial child sexual abuse material websites has doubled, with AI-generated content cited as a contributing factor, the Guardian reported. IWF analysts identified 15,031 commercial CSAM sites in 2025, up from 7,028 in 2024, as criminal gangs profited from the expansion. The IWF is the UK-based organization responsible for identifying and removing CSAM from the internet; its findings are referenced by U.S. and UK policymakers assessing the scale of online child exploitation.

Read at The Guardian ↗