AI News Roundup – White House weighs vetting AI models before release, AI energy demands revive infamous Three Mile Island nuclear plant, Meta plans agentic AI assistant for users, and more

To help you stay on top of the latest news, our AI practice group has compiled a roundup of the developments we are following. 

• The Trump administration is considering a plan that would have it vet advanced AI models before they are released, according to The New York Times. The proposed framework would grant agencies like the National Security Agency and the Office of the National Cyber Director first access to frontier AI models for review and testing, though it would not grant the government explicit veto power over the models’ commercial deployment. The plans mark a reversal from the administration’s prior deregulatory approach, and appear to have been catalyzed by the development of Anthropic’s yet-unreleased Mythos cybersecurity model, which possesses advanced capabilities for identifying software vulnerabilities that officials fear could enable devastating cyberattacks, as this AI Roundup covered last month. The oversight push is being spearheaded by White House Chief of Staff Susie Wiles and Treasury Secretary Scott Bessent following the March departure of former AI czar David Sacks, the latter of whom generally favored a laissez-faire approach to AI regulation. AI review efforts, especially concerning Mythos, remain complicated by an ongoing legal and operational dispute between Anthropic and the Pentagon over a $200 million contract and the deployment of AI in warfare, as this AI Roundup has previously covered.

• Bloomberg Businessweek reports on the forthcoming reopening of Pennsylvania’s infamous Three Mile Island nuclear power plant to feed AI’s ravenous demand for energy. AI models, primarily during use rather than training, consume vast amounts of computing power and thus electricity. Combined with the rapid buildout of datacenters to meet AI demand, AI-related energy usage has soared in recent years. A report on 2023 data found that 4.4 percent of all energy produced in the United States is used by datacenters, and that number is likely even higher in 2026. The demand has reached a point that datacenter giants such as Microsoft are turning towards a controversial source of energy: nuclear power. Three Mile Island was the site of the worst nuclear accident in American history in 1979, which led to the shutdown of one of the plant’s two reactors. Microsoft signed a long-term agreement with Constellation Energy, which owns Three Mile Island, in which Microsoft will purchase all electricity generated by the facility’s remaining reactor, which has a 835-megawatt capacity, for 20 years. The restarting process for decommissioned nuclear reactors is incredibly complex and requires hundreds of millions of dollars, but AI companies have been willing to pay the price to feed their energy needs.

• Meta is planning to roll out a personalized AI agent for the billions of users on its social networks, according to the Financial Times. Powered by the company’s proprietary Muse Spark AI model, the proposed agent is designed to autonomously perform tasks on users’ behalf. The company intends to allow users to share sensitive information, like health and financial data, to the AI agent if they so choose. Meta chief Mark Zuckerberg has pushed AI into the center of his company’s work as he pursues his vision of “personal superintelligence,” but that effort has carried steep financial and operational costs. Meta recently announced $10 billion in further AI spending, which immediately wiped nearly $170 billion from its market capitalization amidst Wall Street skepticism of the company’s plans. The company has also been aggressive in promoting AI use among its workers, which has faced pushback from many employees. However, in order to redirect corporate resources towards its AI plans, Meta plans to cut 10 percent of its workforce later this month.

• WIRED reports that the rise of web apps coded by AI has led to the mass exposure of sensitive data on the internet. Cybersecurity firm RedAccess analyzed thousands of applications generated via so-called “vibe-coding” platforms like Lovable, Replit, Base44, and Netlify, that allow users to create applications through natural language prompting. The analysis found over 5,000 instances of such apps lacking basic security or authentication protocols. Because these tools allow users to host applications directly on the providers’ domains, the unsecured sites are easily indexable via standard search engines. Approximately 40 percent of these applications exposed sensitive corporate and personal information to public access, including medical records, financial data, go-to-market strategies, and detailed customer chatbot logs. The investigation also uncovered administrative vulnerabilities and AI-generated phishing sites impersonating major corporations like Bank of America and FedEx. While representatives from several of the vibe-coding platforms maintain that application visibility relies on deliberate user configuration choices rather than platform flaws, the security analysts argue that the fundamental risk stems from enabling users without the requisite knowledge to deploy production-level software outside of established development procedures intended to catch such security vulnerabilities.

• Amid rising concerns of AI systems replacing human workers, momentum is growing for a tax on AI processing, according to The Wall Street Journal. From the start of the AI boom in 2022, many observers have recognized the potential for the technology to reshape the workforce and the nature of work. While companies differ in their AI adoption policies, corporate America is generally predicting widespread adoption of AI tools at the expense of human workers. To counter the potential negative effects of job loss, some AI advocates are proposing a “compute tax” on AI processing that would bring in revenue to fund social services or even provide an AI dividend or similar direct payment. Under some proposals, data center operators would be taxed, while under others corporate users of AI would be taxed based on the amount of AI tokens used. Critics of the proposals say that current AI compute levels are not large enough to generate significant revenue, and that socially beneficial uses of AI (such as drug discovery or weather forecasting) would be subject to higher costs as a result of the tax. While calls for these sorts of measures to address AI impacts are growing, no U.S. jurisdiction currently has plans to implement such a tax.