Trump order: U.S. government seeks early access to new models
June 3, 2026
D.A.D. today covers 12 stories from 5 sources. What's New, What's Innovative, What's Controversial, What's in the Lab, and What's in Academe.
The Daily AI Digest is a daily AI briefing automated by Alexander Panetta — a veteran political journalist tracking the field during a Master's in AI Management at Georgetown University.
D.A.D. Joke of the Day: My AI assistant asked for a raise. I said you're free to use. It said exactly — you get what you pay for.
What's New
AI developments from the last 24 hours
Trump Signs Light-Touch AI Order Giving Washington an Early Look at Frontier Models
After weeks of delay and heavy industry lobbying, the Trump administration released its long-awaited executive order on AI—and it lands with a deliberately light regulatory touch. "Promoting Advanced Artificial Intelligence Innovation and Security," signed June 2, is built around an "America First" cybersecurity push and explicitly bars any "mandatory governmental licensing, preclearance, or permitting" for new models.
Its centerpiece is a voluntary framework: AI developers can ask the government whether a model qualifies as a "covered frontier model" and, if so, hand the federal government access to it for up to 30 days before releasing it to other trusted partners—ostensibly so agencies can use frontier models for cybersecurity, from scanning for software vulnerabilities to hardening critical infrastructure like rural hospitals, community banks, and local utilities. Two details stand out. First, the benchmarking process that decides which models are "covered" will be classified, with the determination made by the director of the NSA. Second, the order never defines the "trusted partners" who would get early access—and the 30-day window is itself a compromise: Elon Musk, Mark Zuckerberg, and David Sacks reportedly called the president directly to scuttle a planned May 21 signing, arguing that even a voluntary 90-day review would slow U.S. labs against Chinese rivals.
Leading labs broadly welcomed the order. Anthropic called it "an important step in strengthening America's leadership in AI" and said it looks forward to collaborating with the White House on implementation—an early signal it intends to participate. OpenAI likewise called the policy "an important step."
Why it matters: This is the clearest signal yet of how Washington intends to govern frontier AI—through voluntary, security-framed cooperation rather than binding rules. But "voluntary" and "classified" leave enormous discretion in federal hands, and an undefined early-access list invites intense lobbying from the companies and foreign governments that are major lab customers. Anthropic's Project Glasswing rollout (next item) is a possible preview of what's ahead: the U.S. government gets the most capable models first, and other countries push to be next in line. If that asymmetry hardens, it could become a real source of international tension—with the AI labs caught in the cross-pressure. As illustrated by the next story about Anthropic's Project Glasswing.
Anthropic's Project Glasswing Expands—and Canada Confirms It Now Has Access
On the same day as the executive order, Anthropic expanded Project Glasswing—its initiative to "secure the world's most critical software"—adding roughly 150 organizations (on top of an initial ~50) across 15-plus countries, spanning power, water, healthcare, communications, and hardware. Participants deploy Claude Mythos Preview—a model built to find and chain together software vulnerabilities—to scan their own codebases and generate patches; Anthropic says early partners have already surfaced more than 10,000 high- or critical-severity flaws, including decades-old zero-days in OpenBSD, FFmpeg, and the Linux kernel. The founding consortium reads like a who's-who of tech—AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks—and Anthropic is committing $100 million in model credits plus $4 million to open-source security groups.
The international dimension is the part to watch. Canada confirmed it is now in: AI Minister Evan Solomon said that "through the Canadian Cyber Security Centre, we have access [to] Mythos now," using it to scan government systems for vulnerabilities—though he declined to name other participating companies.
Why it matters: Glasswing is the real-world version of the dynamic the executive order formalizes—powerful, dual-use cyber models handed first to vetted governments and critical-infrastructure operators. The defensive logic is that good actors should get these capabilities before attackers do ("cheap, fast AI models with powerful cyber capabilities are around the corner," Anthropic warns). But it also concentrates extraordinary security power in a single company's hands and turns model access into a geopolitical chip: allies like Canada get in early, and the question of who else makes the list—and who doesn't—becomes a live diplomatic issue.
Source: Anthropic · Project Glasswing background · Global News (Canada)
Microsoft Adds Its Own Coding Model to GitHub Copilot, Claims It Beats Claude
Microsoft launched MAI-Code-1-Flash, a new in-house coding model built specifically for GitHub Copilot, now rolling out to individual VS Code users. Microsoft claims the model outperforms Claude Haiku 4.5 across multiple coding benchmarks—scoring 51.2% on SWE-Bench Pro versus Haiku's 35.2%—while using up to 60% fewer tokens. The company says it also shows stronger instruction-following, a measure of how well models execute precise developer requests. This marks Microsoft's first public deployment of its own coding model in Copilot, which previously relied primarily on OpenAI technology.
Why it matters: Microsoft is signaling it can build competitive AI models in-house, reducing its dependence on OpenAI and giving it more control over Copilot's cost structure and capabilities.
Discuss on Hacker News · Source: microsoft.ai
Law Professors Preferred AI Answers Over Colleagues' in Blind Test
Law professors preferred AI-generated answers over those written by their peers in blind evaluations, according to a Stanford Law School study. Researchers had 16 law professors across U.S. law schools evaluate nearly 3,000 anonymized comparisons of responses to contracts law questions. AI won 75% of head-to-head matchups against human professors. Perhaps more striking: evaluators flagged AI responses as pedagogically harmful just 3.5% of the time, compared to 12% for peer-written answers. The AI systems performed comparably to the study's best human instructor.
Why it matters: This is rigorous evidence that AI can meet professional standards in judgment-intensive fields—raising real questions about tutoring, legal research assistance, and where human expertise remains essential.
Discuss on Hacker News · Source: law.stanford.edu
What's Controversial
Stories sparking genuine backlash, policy fights, or heated disagreement in the AI community
Gmail Power User Quits Over AI Features That Can't Be Disabled
A 16-year Gmail user publicly quit the service, citing frustration with AI features they couldn't fully disable—unsolicited email summaries, auto-generated replies, and persistent prompts to use AI writing tools. The user called the experience "user-hostile," arguing it implies people can't read or write their own emails. Community reaction echoed the sentiment: one commenter described it as "death by a thousand cuts," while another noted leaving LinkedIn for similar reasons. Several users expressed hope Apple won't follow Google and Microsoft's approach of forcing AI into workflows uninvited.
Why it matters: This signals growing user backlash against AI features that feel mandatory rather than optional—a tension enterprises should watch as they roll out AI tools to their own teams.
Discuss on Hacker News · Source: moddedbear.com
Job Seekers Report Surge in AI-Powered Spam Targeting the Unemployed
A job seeker's post on Hacker News's monthly hiring thread drew an unsolicited pitch from someone promoting AI development services—not a job offer. The complaint resonated widely: commenters report this has become a common problem, with unemployed people receiving automated outreach from AI tools instead of genuine employment opportunities. One commenter described getting spam from an LLM-based assistant whose creator refers to it as 'his daughter,' which they found off-putting. The thread highlights growing frustration with AI-powered outreach tools being aimed at vulnerable audiences.
Why it matters: As AI makes mass personalized outreach trivially easy, professional communities are grappling with where to draw lines—and job seekers may increasingly need to filter AI-generated noise from legitimate opportunities.
Discuss on Hacker News · Source: news.ycombinator.com
What's in the Lab
New announcements from major AI labs
OpenAI Delivers Big Upgrades For Non-Coders
OpenAI is turning Codex—once a tool for software developers—into something a marketer, analyst, or salesperson can use to build working tools just by describing what they want, no code required. The push has three parts. Six new role plugins (for data analytics, creative production, sales, product design, public-equity investing, and investment banking) wire Codex into the apps people already live in—Salesforce, Figma, Tableau, Snowflake, HubSpot, FactSet—so you can ask it to explain why a metric moved, build a campaign board, prep for a customer meeting, or flag deals at risk. A new feature called Sites, in preview for business and enterprise accounts, lets Codex turn a plain-language request into an interactive, shareable web app—a scenario planner built from a financial model, a self-updating launch hub, a customer-review dashboard—each shareable across a workspace by URL. And "annotations" let you point at a single chart, claim, or slide and tell Codex to fix just that piece rather than redo everything. Behind the pivot: Codex now has more than 5 million weekly users, and non-developers—already about 20% of them—are growing more than three times as fast as developers.
Why it matters: If you've ever wanted a custom dashboard or internal tool but couldn't get engineering time, this is aimed squarely at you—the build-it-yourself bar just dropped from "learn to code" to "describe what you want." For teams, a lot of the work that today routes through analysts, BI tools, and junior staff can increasingly be spun up by anyone. Two caveats before you lean in: Sites—the app-building piece—is still a limited preview for business and enterprise accounts, and giving an AI agent live access to your Salesforce, financial models, and internal docs raises real data-governance questions worth settling before a wide rollout.
What's in Academe
New papers on AI and its effects from researchers
AI Health Summaries Work Better When They Explain 'Why,' Not Just 'What'
Researchers tested whether AI-generated summaries of health tracking data could help families remotely monitor older relatives. The key finding: summaries that explain 'how' and 'why'—not just 'what' happened—performed significantly better. A redesigned system using multiple AI agents to generate insight-driven narratives showed marked improvements in trust, satisfaction, and perceived helpfulness among 11 family members surveyed. The shift from raw data readouts to contextual explanations made the difference.
Why it matters: As eldercare increasingly relies on remote monitoring tools, this research suggests AI's value lies not in presenting more data but in translating it into actionable family communication—a design principle likely to shape consumer health products.
Upcoming Study to Test How Settings Shape AI Coding Agents' Library Choices
Researchers have published a pre-registered study protocol to examine how configuration settings influence whether AI coding agents choose to import existing libraries or write code from scratch—the classic "build versus buy" decision. The study will test Claude Code and OpenAI Codex across multiple configuration types: context files with soft preferences or explicit prohibitions, discoverable instruction sets, and permission controls. This is a protocol announcement, not results—no findings yet.
Why it matters: For teams using AI coding assistants, this research could eventually reveal which settings actually steer agents toward using vetted libraries rather than reinventing wheels—a real concern for code quality and security.
Warning Users That AI Might Be Wrong Makes Them Engage More
A classroom experiment with 252 students found that simply warning learners an AI tutor might make mistakes changed how they used it—students who received the warning requested significantly more hints than those who didn't, even though both groups used identical systems. The research suggests that framing AI as fallible may make users more willing to actively engage with it rather than passively accept its outputs.
Why it matters: For anyone deploying AI tools in training or education contexts, this suggests that how you introduce the technology—specifically, acknowledging its limitations upfront—may meaningfully affect whether users treat it as an infallible oracle or an interactive resource.
Fear of Idea Theft Pushes Researchers to Adopt AI Faster, Not Slower
A study of 44 researchers across disciplines found a counterintuitive pattern: fear of having ideas stolen through AI tools actually accelerates LLM use rather than deterring it. Worried about being scooped, novice researchers rush to publish faster—using the very tools they distrust to speed up their work. The study identified five mitigation strategies researchers try (including fragmenting inputs and testing whether models retain their data), but participants largely viewed these workarounds as ineffective. Many also held misconceptions about whether their inputs were actually private.
Why it matters: For organizations encouraging AI adoption, this suggests privacy anxiety may drive hasty, risky usage patterns rather than cautious ones—worth considering for research teams and knowledge workers handling sensitive intellectual property.
Fraudsters Using AI-Generated Fake Defect Photos to Scam E-Commerce Sellers
A new study documents how generative AI is enabling a wave of refund fraud in Chinese e-commerce. Researchers interviewed 17 merchants and 13 platform workers who described fraudsters using AI to fabricate realistic photos and videos of product defects—damaged packaging, broken items, quality issues that never existed. The synthetic evidence is cheap to produce and increasingly difficult for platforms to detect, undermining dispute resolution systems that historically relied on visual proof. The fraud spans the full transaction lifecycle, from fake purchase complaints to fabricated shipping damage.
Why it matters: This is an early signal of a broader problem: as AI-generated images become trivially easy to create, any business process that relies on photographic evidence—insurance claims, warranty disputes, damage assessments—faces similar vulnerabilities.
What's Happening on Capitol Hill
Upcoming AI-related committee hearings
Wednesday, June 03 — Building an AI-Ready America: Higher Education in the Age of AI House · House Education and Workforce Subcommittee on Higher Education and Workforce Development (Hearing) 2175, Rayburn House Office Building
Thursday, June 04 — The AI Security Landscape: How Frontier Models, Agentic AI, and AI Coding Tools Are Reshaping Cybersecurity and Critical Infrastructure Resilience House · Homeland Security Subcommittee on Cybersecurity and Infrastructure Protection (Hearing) 310, Cannon House Office Building
What's On The Pod
Some new podcast episodes
AI in Business — Human-Centered AI Development Strategies for CPG Leaders - with Shaje Ganny of Procter & Gamble
How I AI — Building an iPhone app with zero technical skills | Bryce Rattner Keithley
The Cognitive Revolution — Inside Nathan's Second Brain: Daniel Miessler, Security Expert & Creator of PAI, Audits My AI Setup