AI Briefing

Washington's grip on frontier AI tightened again this week: after forcing Anthropic to pull its most powerful models, the White House has now asked OpenAI to release GPT-5.6 customer-by-customer to government-approved partners only. Wall Street is getting nervous too, with Big Tech shedding a staggering $2.7 trillion in June as investors finally ask who pays for all this compute. Underneath the drama, the plumbing of the agentic web is being laid (Google's new ARD standard, Amazon's AgentCore stack), and a quietly humbling study shows today's smartest models flunk a test most schoolchildren can pass.

01PolicyCNN ↗

White House Tells OpenAI to Stagger GPT-5.6's Release, Approving Customers One by One

A source told CNN the White House has asked OpenAI to limit its upcoming GPT-5.6 model to a small set of government-approved partners, with access granted 'customer by customer,' because officials view it as 'on par' with Anthropic's restricted Mythos model. It marks the second time this month Washington has intervened in a frontier launch, after the Commerce Department's export-control order forced Anthropic to pull Mythos and Fable. Sam Altman told staff this staggered approach is 'not our preferred long term model,' underscoring that the U.S. still has no clear, consistent framework for who regulates advanced AI or how.

02Open SourceGoogle Developers Blog ↗

Google and 10 Rivals Publish ARD, an Open Standard So AI Agents Can Find and Trust Each Other

Google released the Agentic Resource Discovery (ARD) specification, an open, Apache-2.0-licensed protocol that lets AI agents publish, discover, and cryptographically verify tools, skills, and other agents across organizational boundaries. Launch partners include Microsoft, Nvidia, Salesforce, Snowflake, GitHub, Hugging Face, Databricks, Cisco, ServiceNow and GoDaddy. It works through two primitives: an 'ai-catalog.json' manifest hosted on a company's own domain (so domain ownership becomes the root of trust) and registries that crawl those catalogs and answer plain-language discovery queries. ARD aims to be the missing 'search and trust' layer of the agentic web, the same role DNS and search engines played for the open internet.

03ProductAWS News Blog ↗

Amazon Makes AgentCore Generally Available, Adding Managed Knowledge and Web Search for Production AI Agents

At its New York Summit, AWS moved its agent stack from preview to production: the Bedrock AgentCore harness is now generally available, letting developers build and run production-grade agents purely from configuration without writing orchestration loops. AWS also launched a managed Knowledge Base with an agentic retriever, a fully managed Web Search tool that grounds agents in cited web results with zero data egress, and a forthcoming 'AWS Context' service that maps company data into a knowledge graph. It even rolled out a WAF capability that lets publishers charge AI bots for content access, a sign the agent economy is starting to grow its own toll booths.

04ResearchScienceDaily (PNAS Nexus) ↗

A Classic Psychology Test Just Exposed a Deep Flaw in Today's Best AI Models

Researchers led by Suketu Patel ran leading models including GPT-5, Claude Opus 4.1 and Gemini 2.5 through the Stroop task, a decades-old attention test where you must name the ink color of a word rather than read the word itself. The models did fine on short lists but collapsed as lists grew: GPT-4o fell from 91% accuracy on five words to just 15% on forty, and accuracy on conflicting items dropped toward zero as the systems defaulted to simply reading the words. Published in PNAS Nexus, the study argues this points to a fundamental gap between transformer 'attention' and the executive control humans use to stay focused under distraction.

05FundingStartup Fortune ↗

Big Tech Sheds $2.7 Trillion in June as Investors Ask Whether AI Can Pay for Itself

The 'Magnificent Seven' plus Broadcom and Oracle have lost roughly $2.7 trillion in market value this month, per a Yahoo Finance analysis, as Wall Street reprices the AI build-out as a cost rather than a pure growth story. Alphabet alone shed about $225 billion in a single session on June 22, its largest one-day wipeout on record, amid reports of up to $190 billion in planned 2026 AI spending and the high-profile departures of researchers Noam Shazeer and John Jumper. With the four biggest hyperscalers on track to spend roughly $725 billion in capex this year, investors are now watching free cash flow as closely as model demos.