Today's Headlines

  • The Fable 5 and Mythos 5 shutdown reportedly began with concerns Amazon's CEO raised to the government
  • KPMG pulls its "age of agentic AI" report after AI-generated inaccuracies surface
  • China's MiniMax open-sources M3, a 428-billion-parameter model that matches Gemini 3.1 Pro

Yesterday's export control order entered a new phase today.

New reporting surfaced on what set yesterday's Fable 5 and Mythos 5 shutdown in motion.

Separately, a report on AI usage that was itself written with AI was withdrawn after it turned out to contain false claims.

On the performance front, a new open model arrived that stands shoulder to shoulder with Gemini 3.1 Pro.

Today's Top Three

The Fable 5 shutdown reportedly began with concerns Amazon's CEO raised to the government

New reporting has surfaced on what set yesterday's Fable 5 and Mythos 5 shutdown in motion.

According to the Wall Street Journal, Amazon CEO Andy Jassy told Treasury Secretary Scott Bessent and other government officials that Amazon researchers had used Claude Fable 5 to obtain information useful for cyberattacks. That disclosure reportedly preceded the U.S. government's export control directive against the two models.

David Sacks, who led AI policy in the prior Trump administration, separately said that a "highly credible trusted partner" had presented evidence of a jailbreak vulnerability, and claimed the administration asked Anthropic CEO Dario Amodei to fix the issue or remove the model, but that "Dario refused." Anthropic disputes this account of events.

An Amazon spokesperson said it is "not uncommon for governments to seek our counsel on potential security risks," but declined to discuss the specifics of those conversations.

As compensation for the suspension, Anthropic reportedly reset usage limits for all Pro and Max plan users. Access to models other than Fable 5 and Mythos 5 remains available, and the company says it is working to restore access.

The directive itself — the receipt time, scope, and its standing as a precedent — was covered in yesterday's extra edition. What emerged today is the reporting on how the shutdown came about.

Source: Amazon CEO reportedly raised Anthropic model concerns before government crackdown (TechCrunch) / Amazon security research reportedly led to the White House's Anthropic Fable ban (The Verge)

KPMG pulls its "age of agentic AI" report

KPMG has withdrawn a report on AI usage.

The withdrawn document, titled "Redefining excellence in the age of agentic AI," was published in October 2025. GPTZero, a research firm that builds AI-detection tools, found numerous inaccuracies and told the Financial Times that they stemmed from AI hallucinations — content generated without basis in fact.

UBS, the UK's National Health Service, Swiss Federal Railways, and Transport for London each stated that the report's claims about their AI usage were untrue or misleading. In an ironic turn, a report about putting AI to work contained AI-generated falsehoods.

A KPMG spokesperson said the firm removed the report from its website while investigating, adding that "we expect all our people to follow our guidelines on the responsible use of AI, including human oversight to validate content and verify independent sources."

A similar case occurred last month at EY, where a report containing fake footnotes was withdrawn — a recurring pattern at large consulting firms.

Source: KPMG pulls report on AI usage due to apparent hallucinations (TechCrunch)

China's MiniMax open-sources M3, a 428-billion-parameter model matching Gemini 3.1 Pro

China's MiniMax has released a large language model called M3.

It uses a mixture-of-experts (MoE) architecture with roughly 428 billion total parameters and about 23 billion active. A proprietary technique called MiniMax Sparse Attention enables a context window of up to 1 million tokens, with prefill speed reportedly about 9 times faster and decode speed about 15 times faster than the previous generation.

On benchmarks, M3 scored 59.0% on SWE-Bench Pro (versus Gemini 3.1 Pro's 54.2%) and 74.2% on MCP Atlas (versus 69.2%), edging out Gemini 3.1 Pro. It fell slightly behind on BrowseComp (83.5% to 85.9%) and OSWorld-Verified (75.2% to 76.2%), leaving the two roughly even overall.

The model weights were published on Hugging Face on June 12 and are free for non-commercial use. Individuals and companies with annual revenue under $20 million may use the model commercially after registering with MiniMax and crediting "Build with MiniMax."

Source: MiniMax M3, a 428B-parameter LLM matching Gemini 3.1 Pro, goes open (PC Watch) (Japanese)

Other Developments

Models & Products

  • Z.ai released GLM-5.2, a coding-focused LLM supporting up to 1 million tokens of context and up to 128K output tokens. It is currently available to GLM Coding Plan subscribers, with MIT-licensed weights and a metered public API planned for next week (PC Watch, Japanese)
  • Apple's new AI photo editing features for iOS and macOS — including Reframe, Extend, and Clean Up — proved useful but inconsistent in hands-on testing, with quality varying by feature (The Verge)

Research

  • Developer Koen van Gilst documented using Claude Fable to generate a playable browser game, "Shepherd's Dog," in about 45 minutes as a single self-contained HTML file, noting that earlier Claude models had failed at the same task (koenvangilst.nl)

Business

  • OpenAI received a subpoena from a coalition of state attorneys general led by New York, seeking documents on advertising practices, user engagement, data handling, and protection of minors and seniors; the company said it would engage constructively with the investigation (TechCrunch)
  • Entrepreneur Andrew Yang argued that the next major startup opportunity lies in businesses that lower consumer costs in areas like housing, education, food, and wireless service; he has launched Noble Mobile, a low-cost carrier that returns unused data costs to customers (TechCrunch)
  • Meta drew sharp criticism from employees after reassigning roughly 6,500 people to its three-month-old Applied AI unit for work such as generating model training data; more than 1,600 employees signed a petition protesting keystroke monitoring (TechCrunch)

Policy & Safety

  • A Derbyshire Police officer is under investigation for allegedly using AI to fabricate or alter evidence across multiple cases, raising governance questions about AI use in criminal proceedings (Sky News)

Other

  • TensorZero, an open-source LLMOps platform, archived its GitHub repository as read-only shortly after announcing a $7.3 million seed round, prompting debate over the abrupt closure following funding (GitHub)
  • Paca, a self-hosted project management tool that embeds AI agents as full Scrum teammates on shared boards and connects to Claude and other agents via MCP, was released (GitHub)

Sources: AI news inbox collected June 14, 2026 (27 items; 1 primary, 26 secondary), curated by the editorial desk.