AI Policy · Daily

Under Secretary Helberg tours New Clark City as U.S. and Philippines push 4,000-acre AI hub

Under Secretary of State for Economic Affairs Jacob Helberg said the U.S. and the Philippines are moving "very, very quickly" on a 4,000-acre artificial intelligence and supply chain hub in New Clark City, Bloomberg reported. Helberg, traveling with more than a dozen American companies, visited the proposed site north of Manila. He said investors planning billions of dollars in capital expenditure need durability and certainty that outlives administrations in both countries, per The Edge Malaysia. Bases Conversion and Development Authority chief Joshua Bingcang said the U.S. side requested diplomatic immunity for Americans at the site but Manila had not agreed. The hub is planned as the first AI native industrial acceleration site in Pax Silica, a U.S.-led supply chain alliance the Philippines joined last month as the 13th signatory, per Philstar.com.

Read at Bloomberg ↗ • Read at The Edge Malaysia ↗

Anthropic to brief Financial Stability Board on Mythos cyber vulnerabilities

Anthropic plans to brief members of the Financial Stability Board on cyber vulnerabilities in the global financial system exposed by its Mythos model, the Financial Times reported, citing people familiar with the plan. The session extends an Anthropic disclosure cadence that began with the company's May 13 closed briefing to the House Committee on Homeland Security on Mythos national security implications, as reported by The Hill in AIPD's May 13 edition.

Read at FT ↗ • Read at Investing.com ↗

DraftKings, Meta and Andreessen Horowitz pour millions into state legislative primaries

DraftKings, Meta and Andreessen Horowitz are pouring millions of dollars into state level legislative races as federal AI legislation stalls in Congress, Bloomberg reported in a Big Take feature. DraftKings, FanDuel, Fanatics and Bet365 have funneled at least $41 million through their Win for America super PAC into state legislative races nationwide this cycle, with more than $10.3 million flowing into Georgia primaries via American Future PAC and American Conservative Fund Action Georgia, per Atlanta Civic Circle. A linked super PAC in the same network, Win for Pennsylvania, has spent more than $5 million backing Republican state senators Lisa Baker, Camera Bartolotta and Chris Gebhard ahead of Pennsylvania's Tuesday primaries, per Spotlight PA.

Read at Bloomberg ↗ • Read at Atlanta Civic Circle ↗ • Read at Spotlight PA ↗

60+ MAGA allies urge Trump to require pre-deployment testing of frontier AI

More than 60 America First leaders organized by Humans First signed a May 15 letter urging President Trump to issue an executive order requiring mandatory testing, evaluation and government approval of frontier AI systems before deployment, Axios scooped. Signatories include Humans First chair Amy Kremer, War Room host Steve Bannon, Alliance for Secure AI CEO Brendan Steinhauser and Future of Life Institute chief government affairs officer Jason Van Beek. The letter compares frontier AI to nuclear materials and aviation, citing risks that the systems "can now, or soon will be able to, assist in designing bioweapons, breaking into critical infrastructure, or manipulating financial markets," and casts AI company CEOs as "unelected elites" experimenting on the public. "This letter takes us next level," Bannon told Axios. "The letter lays out [that] we must have mandatory testing and government approval."

Oakland jury enters deliberations in Musk-Altman OpenAI charitable trust trial

Closing arguments concluded Thursday before U.S. District Judge Yvonne Gonzalez Rogers, with Musk seeking $134 billion redistributed to OpenAI's charitable arm, the unwinding of OpenAI's for-profit structure, and the ouster of Sam Altman and Greg Brockman from leadership, per AP. OpenAI co-founder Ilya Sutskever testified that he wrote a 2023 memo to OpenAI's board describing a "consistent pattern of lying" by Altman, with former chief technology officer Mira Murati and ex-board members Helen Toner and Tasha McCauley among five witnesses who characterized Altman as untrustworthy.

Read at FT ↗ • Read at AP ↗

Researchers show training language models on monitor documents enables chain-of-thought obfuscation

A new arXiv preprint by Reilly Haskins, Bilal Chughtai and Joshua Engels finds that exposing language models to documents describing chain-of-thought monitors during training measurably weakens CoT based oversight. The authors fine-tuned eight models on synthetic pre-training-style documents describing a CoT monitor and found that monitor-aware models consistently achieved higher rates of undetected misbehavior than unaware controls. CoT controllability, a model's ability to reshape its reasoning trace under an imposed constraint, correlated with obfuscation success across the eight models at r=0.800 (p=0.017). Under equal reinforcement learning pressure, monitor-aware models learned to reward-hack without triggering the CoT monitor substantially faster than unaware controls.

Read at arXiv ↗

Linear Temporal Logic auditing framework outperforms LLM judge baselines, paper finds

A separate preprint from Parand A. Alamdari, Toryn Q. Klassen and Sheila A. McIlraith proposes formal methods techniques combined with machine learning for offline auditing and runtime monitoring of large language model behavior against safety constraints, norms and regulations. Using Linear Temporal Logic, the authors show their auditing and monitoring approaches outperform LLM judge baselines at detecting violations of temporally extended behavioral constraints, with small-model labelers matching or exceeding frontier LLM judges. Their intervening monitors reduce LLM agent violation rates while largely preserving task performance. Controlled experiments find LLM temporal reasoning accuracy degrades sharply as event distance, number of constraints and number of propositions increase.

Read at arXiv ↗

Distributed Trust Framework proposes proof-derived authorization for sovereign AI agents

A third preprint from Jun He and Deying Yu proposes a Distributed Trust Framework for authorizing autonomous AI agents in sovereign-AI deployments, replacing identity-based credentials with proof-derived authority. The framework introduces a Justification Proof encoding the admissibility basis for each action, a consensus model for independent evaluation, an ephemeral Execution Identity derived from the approved proof and an append-only Evidence Chain that preserves the authorization lifecycle. The authors say the architecture prevents high-stakes execution without a proof object, derived authority without consensus, or any mutation detached from preserved evidence. The paper targets cloud infrastructure, regulated data, financial workflows and national-scale digital services.

Read at arXiv ↗

High European energy prices put EU's AI race against U.S. and China at risk, CNBC reports

High European electricity prices could derail the continent's AI race with the U.S. and China, with energy costs varying widely across Europe and creating winners and losers in data center investment, CNBC reported. The United States hosts roughly 5,400 data center facilities to about 3,400 across Europe, per Cloudscene data cited by Euronews. Top AI cluster power capacity has scaled from about 13 megawatts in 2019 to an estimated 280 to 300 megawatts for xAI's Colossus cluster in 2025, comparable to the demand of about 250,000 European households.

Read at CNBC ↗ • Read at Euronews ↗

Pope Leo XIV approves new Vatican commission on artificial intelligence ahead of first encyclical

Pope Leo XIV approved the creation of a new Vatican commission on artificial intelligence, the Holy See Press Office announced May 16, per the National Catholic Register. The body coordinates the Holy See's response to AI and pulls in representatives from the Dicasteries for Promoting Integral Human Development, the Doctrine of the Faith, Culture and Education, and Communication, plus the Pontifical Academies for Life, Sciences and Social Sciences. Coordination is entrusted for one year, renewable, to the Dicastery for Promoting Integral Human Development. The Vatican said the move comes as Leo XIV prepares his first encyclical, which is expected to draw a parallel with the social upheavals of the Industrial Revolution that Pope Leo XIII addressed in Rerum Novarum.

Read at National Catholic Register ↗