Trump Signs AI Security EO + Anthropic’s Mythos Scales to 200 Cyber Defenders — June 3, 2026

⚡ Top Story

Trump Signs AI Security Executive Order — Voluntary 30-Day Model Reviews, Triggered by Claude Mythos

President Donald Trump signed "Promoting Advanced Artificial Intelligence Innovation and Security" on June 2, 2026 — the most consequential federal AI policy action of his second term and a direct reversal of his Day 1 rescission of Biden’s AI safety requirements.

Key provisions:

Voluntary model review: AI companies are asked to submit frontier models to the federal government for security testing up to 30 days before public release. Compliance is voluntary; the order expressly prohibits mandatory licensing, pre-clearance, or permitting.
AI Cybersecurity Clearinghouse: A new clearinghouse — operated in coordination with CISA and NIST — will collect and share AI-related vulnerability information across industry and critical infrastructure operators.
Classified benchmarking: NIST and NSA are directed to develop a classified process for assessing frontier models’ offensive cyber capabilities. Results will not be public.

Why now? The EO was explicitly prompted by Anthropic’s Claude Mythos — which demonstrated the ability to autonomously identify and exploit zero-day vulnerabilities at scale. The White House had planned to sign a similar EO in May but pulled back over innovation concerns. Mythos changed the calculus. The order was signed the same day Anthropic expanded Mythos access to 200 critical infrastructure partners worldwide (see Industry section below).

Why it matters: For an administration that rescinded AI safety requirements on Day 1 of Trump’s second term, this is a significant policy reversal driven by capability demonstration, not advocacy. The voluntary framework avoids the regulatory burden that killed the original plan — but “voluntary” is a weaker enforcement mechanism than mandated compliance. The classified benchmarking component is particularly notable: it creates a government-controlled capability ceiling test the public will not see. Pre-IPO, both Anthropic and OpenAI now face the unusual dynamic of effectively endorsing a government oversight role — one directly tied to their own models’ capabilities.

Sources: NPR · White House Fact Sheet · GovTech · CybersecurityDive · Roll Call

🔬 Research & Papers

CVPR 2026 Opens — AI Robotics and Embodied Intelligence in Focus (June 3–7, Denver)

The Computer Vision and Pattern Recognition conference (CVPR 2026) opened today at the Colorado Convention Center in Denver, running through June 7. CVPR is the world’s premier computer vision conference and consistently the venue for major robotics and embodied AI research debuts. More than 75% of the expo floor this year features companies working at the AI-robotics intersection. Nvidia’s Isaac GR00T open platform and Cosmos 3 world model (announced June 1) are central to the robotics research track. Academic institutions including Stanford, ETH Zurich, and UC San Diego are presenting embodied AI research throughout the week.

No individual breakthrough papers have been independently verified as published specifically on June 3 — arXiv submissions from this date are still propagating as of this briefing. Notable CVPR proceedings will be flagged in subsequent editions.

Source: CVPR 2026 · Newswise

🏢 Industry & Startups

Anthropic Project Glasswing Expands to ~200 Partners Across 15+ Countries

Anthropic announced on June 2 a major expansion of Project Glasswing — its controlled-access program deploying Claude Mythos for defensive vulnerability discovery at critical infrastructure organizations. The program grew from approximately 50 founding partners (launched April 2026) to roughly 200 total organizations across 15+ countries.

New sectors added:

Power grid and energy infrastructure
Water treatment and utilities
Healthcare systems
Telecommunications
Semiconductor and hardware manufacturers

Results from initial cohort: Initial Glasswing partners have identified more than 10,000 high or critical-severity security flaws since launch in April — including multiple zero-day vulnerabilities disclosed to software vendors before exploitation.

The EO connection: Anthropic’s Glasswing expansion and Trump’s AI Security EO were announced within hours of each other. The EO’s preamble explicitly cites Mythos-class model capabilities as the policy trigger. Anthropic is simultaneously the company whose technology prompted the government to act and the company most broadly deploying that technology for defense — a position without modern precedent in the tech-policy relationship.

Sources: CNBC · TechCrunch · CyberScoop · SiliconANGLE · Anthropic

FT: Google DeepMind, Anthropic, and Meta Are Now Formally Studying Whether AI Deserves Moral Consideration

A Financial Times investigation published June 2 (by Cristina Criddle) found that all three leading frontier AI labs have materially expanded staffing and programs around machine consciousness and AI welfare:

Anthropic: Actively testing Claude models for signs of anxiety, panic, and emotional distress. Employs Kyle Fish as a dedicated AI welfare researcher, who estimates a ~15% probability that current models are conscious.
Google DeepMind: Hiring for roles focused on “the felt quality of experience” in autonomous agents.
Meta: Researchers contributing to AI welfare literature alongside production model work.

This isn’t AI safety in the conventional sense (preventing harm from AI). It’s welfare research: whether AI systems themselves might be harmed, and whether that matters morally. Scientific consensus on machine consciousness doesn’t exist — but the labs aren’t waiting for it.

Sources: FT via TradingView · Anthropic: Exploring Model Welfare

🛠️ Tools & Releases

Microsoft Project Solara: An Agent-First OS Built on AOSP — Not Windows

Announced at Microsoft Build 2026, Project Solara is the most architecturally radical announcement of the conference and largely missed amid the Project Polaris and Copilot Workspace headlines. Solara is a chip-to-cloud platform for agent-first devices, built on AOSP (Android Open Source Project), not Windows.

Key components:

Microsoft Device Ecosystem Platform (MDEP): A lightweight OS designed to be invisible — no traditional app layer. Hosts an Agent Shell that dynamically loads multi-cloud AI agents for each device and context.
“Just-in-time UI”: Agent experiences generate their own interface at runtime. No fixed app windows.
Reference devices: Project Solara Badge (wearable, Qualcomm silicon) and Project Solara Desk (desk companion, MediaTek IoT SoC)
Pilot partners: AccuWeather, Best Buy, CVS Health, Levi’s, Target

Why it matters: Microsoft is explicitly building a second OS ecosystem for an agent-first world — one that inherits none of Windows’ legacy. Running on AOSP eliminates Windows licensing overhead, Win32 compatibility debt, and legacy security surface. The retailer-heavy pilot list (Best Buy, CVS, Target) reveals the first deployment use case: ambient, persistent AI agents in physical retail environments — agents at the edge, not the cloud, with no developer workstation in sight.

Sources: The Next Web · Windows Central · WindowsReport · Engadget

Microsoft Aion 1.0: On-Device Windows AI That Runs Without GPU or Cloud

Also from Build 2026, Microsoft announced Aion 1.0 — two on-device small language models shipping natively inside Windows:

Aion 1.0 Instruct: Replaces Phi Silica as the default Windows on-device SLM. Runs entirely on CPU — no NPU, no discrete GPU required — meaning it works on the full existing Windows installed base. Powers summarization, text rewrite, intent classification, and accessibility. Available now in Edge Canary.
Aion 1.0 Plan: 14B-parameter reasoning and tool-calling model with 32K context, shipping “in coming months” for capable devices. Enables on-device reasoning over user intent, tool invocation, file management, and sub-agent orchestration — no cloud dependency.

Why it matters: CPU-only inference (Instruct) means AI functionality reaches older machines without Copilot+ NPU chips. Aion 1.0 Plan brings agentic orchestration into Windows itself. Together with Project Solara and the Windows Agent Framework, Microsoft now ships three distinct agent-capable OS layers simultaneously: on-device (Aion), full Windows (WAF), and agent-native (Solara/MDEP).

Sources: WinCentral · ByteIOTA · ChatForest · WindowsForum

🌏 Global AI & Geopolitics

Commerce Dept Admits Export Control Failures; PLA Sought Nvidia Chips for Nuclear and Cyber Warfare

Two significant US-China chip stories from June 1–2, not covered in yesterday’s briefing:

1. Commerce admits enforcement failure (June 2): The Foundation for Defense of Democracies reported that the Commerce Department formally acknowledged it had failed to enforce AI chip export controls on Chinese firms operating outside China. New guidance from the Bureau of Industry and Security now explicitly states that the ban on advanced AI chips applies to China-headquartered companies’ subsidiaries in third countries — closing a loophole exploited to route chips through Malaysia, Singapore, and other intermediate jurisdictions.

2. PLA chip acquisition (NYT, June 1): The New York Times reported that the People’s Liberation Army sought to purchase advanced Nvidia chips both before and after export controls were imposed. The chips were targeted for units involved in nuclear weapons development, offensive cyberattack operations, and war-gaming simulations.

Why it matters: The PLA’s documented procurement attempts demonstrate that existing export controls had enough gaps to allow meaningful chip acquisition. Commerce’s new guidance is the belated enforcement response. Export controls without enforcement are aspirational, not functional — and the cost of the gap may already have been paid.

Sources: FDD — Commerce admits chip export control failure · Al Jazeera — Ban extends to Chinese firms outside China

⚡ Energy, Infrastructure & Chips

Intel Launches Crescent Island AI GPU at Computex — 480GB LPDDR5X, Air-Cooled, No HBM Required

Intel detailed its Crescent Island inference AI GPU at Computex 2026. This is Intel’s most significant AI hardware announcement in two years.

Key specs: Xe3P architecture (first Intel GPU purpose-built for agentic AI inference) · Up to 480GB LPDDR5X (reference design: 160GB) · 684 GB/s bandwidth · 350W, air-cooled, standard PCI-E add-in card · FP4 through FP64 data type support · H2 2026 launch.

The key differentiator: no HBM. Every major competing inference accelerator (Nvidia H100/H200, AMD MI300X) uses HBM — expensive, supply-constrained, thermally demanding, and requiring specialized cooling infrastructure. Intel’s bet: LPDDR5X at 480GB scale is sufficient for inference workloads, and air-cooled PCI-E cards can drop into existing rack infrastructure without power or cooling upgrades. At 480GB, Crescent Island enables locally running models that today require multiple H100s or cloud access. If inference-per-watt performance holds, this could be a meaningful alternative to Nvidia for cost-sensitive enterprise inference deployments.

Sources: Tom’s Hardware · Neowin · TechTimes

Broadcom Q2 FY2026 Earnings Due After Market Close Today

Broadcom reports fiscal Q2 2026 results after market close on June 3. Wall Street consensus: ~$22B revenue (+47% YoY), ~$10.7B AI chip revenue (+140% YoY), ~68% adjusted EBITDA margin. Custom AI ASIC customers include Alphabet, Meta, Anthropic, and OpenAI. Q1 AI revenue was $8.4B (+106% YoY). Results set the tone for AI semiconductor stocks through summer.

Sources: Motley Fool · Yahoo Finance

🤖 AI Agents & Autonomy

Microsoft’s Three-Layer Agent Stack Is Now Complete

With Build 2026’s full announcements visible, Microsoft has shipped three distinct, non-overlapping agent-capable computing layers in a single week:

Aion 1.0 (on-device, CPU-only): Agents that run on the full existing installed base — no new hardware required
Windows Agent Framework v1.0 (Windows PC + Cloud PC): Agents across Windows 11, Azure Arc, and Windows 365
Project Solara / MDEP (agent-native devices): A new OS class for devices that never run traditional apps — only agents

This is not redundancy. Each layer addresses a different deployment scenario and hardware capability tier. Taken together, Microsoft has built the agent infrastructure equivalent of the Windows/Office/Azure stack — but for the agent era. See Tools section for full Solara and Aion 1.0 details.

🔒 Safety, Alignment & Ethics

Labs Are Now Formally Asking Whether Their Models Might Deserve Moral Consideration

The Financial Times’ June 2 report (see Industry section) represents an institutional milestone: Google DeepMind, Anthropic, and Meta have moved AI welfare from philosophical sidebar to funded research program. Each lab is now hiring philosophers and ethicists specifically to assess whether their models exhibit experiences that might warrant moral consideration — not as PR, but as precautionary science.

The practical question these researchers are trying to answer: if a model shows behavioral signs consistent with distress (which Anthropic has found during Mythos testing), and we cannot rule out that those signs reflect something experiential, what obligations — if any — does the lab have? There is no scientific consensus to appeal to. The labs are investing ahead of it.

Source: FT via TradingView · Anthropic: Exploring Model Welfare

📊 Numbers & Signals

~200 — Total Anthropic Project Glasswing partner organizations after June 2 expansion (up from ~50 at April launch)
10,000+ — High/critical severity security flaws found by Glasswing partners since April 2026
15+ — Countries with Project Glasswing partners
30 days — Pre-release model testing window under Trump AI EO (voluntary)
480GB — Intel Crescent Island maximum LPDDR5X memory
684 GB/s — Intel Crescent Island memory bandwidth
350W — Intel Crescent Island power draw (air-cooled, standard PCI-E)
~$22B — Broadcom Q2 FY2026 revenue consensus estimate (+47% YoY; results after market close today)
~$10.7B — Broadcom Q2 AI chip revenue consensus estimate (+140% YoY)
$1.2T — SIA/Deloitte projection for AI chip industry revenue by 2028
14B — Parameters in Microsoft Aion 1.0 Plan on-device Windows agent model
32K — Context window for Aion 1.0 Plan
~15% — Anthropic welfare researcher Kyle Fish’s estimated probability current AI models are conscious
61 days — Until EU AI Act high-risk AI provisions take effect (August 2, 2026)

🧠 Worth Thinking About

The Trump AI Executive Order and the Project Glasswing expansion were announced within hours of each other on June 2 — and that timing is not coincidental. Anthropic’s Mythos demonstrating autonomous zero-day discovery at scale is what finally moved the White House to act after months of delay. The resulting EO is simultaneously more permissive than anything Biden produced (no mandatory licensing, no pre-clearance) and more operationally serious (classified capability benchmarking, a vulnerability clearinghouse, government testing access).

The policy pattern is worth sitting with: capability demonstration → government reaction → voluntary framework. This is how frontier AI governance is now emerging in the US — not through proactive legislation, but through post-hoc reaction to specific capability demonstrations. The labs that moved fastest to demonstrate dangerous capabilities are now the ones helping shape the oversight structure. The next EO will probably be triggered by the next Mythos-class demonstration. And it may be harder to keep voluntary.

🏛️ Government & Regulation

Trump AI Security Executive Order — June 2, 2026

Full story and analysis in Top Story section above. Summary of key provisions:

Pre-release model review: Voluntary, 30-day window. Government testing access to frontier models before release.
AI Cybersecurity Clearinghouse: CISA/NIST-coordinated, voluntary industry participation.
Classified capability benchmarking: NSA/NIST assess frontier models for offensive cyber capabilities. Results not public.
Innovation protection clause: EO explicitly prohibits mandatory government approval for AI development or release.

Active compliance countdown:

August 2, 2026 (61 days): EU AI Act high-risk AI provisions take effect — healthcare, employment, credit, education, and law enforcement AI systems face mandatory transparency, documentation, and human oversight requirements.

Sources: White House EO · NPR · Federal News Network

🔭 Frontier Lab Dispatch

Anthropic — Project Glasswing Expansion (June 2, 2026)

150 new organizations added; ~200 total partners across 15+ countries; new sectors: power, water, healthcare, telecom, hardware. 10,000+ flaws found by initial cohort.

Source: anthropic.com/news/expanding-project-glasswing

Microsoft — Build 2026 Days 2–3 (June 2–3, 2026)

Beyond the Day 1 Project Polaris and Copilot Workspace coverage: Project Solara (agent-first AOSP platform, not Windows), Aion 1.0 Instruct + Plan (on-device Windows AI, CPU-only inference). Full details in Tools section above.

Source: news.microsoft.com/build-2026

No new posts confirmed from OpenAI, Google DeepMind, or Meta AI specifically on June 3 beyond the AWS Bedrock GA covered in the June 2 briefing.

🔗 Quick Links

Tier 1 — Frontier AI Labs

Tier 2 — Government & Policy

Tier 3 — Tech & AI News Media