AI Weekly #24/2026: AI Emergency Shutdown – US Government Halts Anthropic's Fable 5

TL;DR

This week in 30 seconds:

• AI Emergency Shutdown: The US government halts Anthropic’s Fable 5 & Mythos 5 via directive – three days after launch, worldwide, effective immediately.
• Siri Becomes Gemini: Apple replaces Siri’s AI engine with Google Gemini; iOS 27 runs from iPhone 11 onward – EU and China remain locked out due to regulations.
• IPO + Attorney General: OpenAI files an S-1 with the SEC, projecting an $85B loss for 2028 – while simultaneously receiving a subpoena from US state attorneys general.
• Prometheus Awakens: Jeff Bezos launches AI startup with $12B in capital and a $41B valuation, 150 employees – goal: develop jet engines and pharmaceuticals via AI.

---

Audio Version

14:20 | Download MP3

Chapters

- 0:00 - TL;DR - 0:54 - Story of the Week - 3:49 - More Top Stories - 9:25 - Tool of the Week - 10:28 - Fail of the Week - 11:52 - Number of the Week - 12:40 - Reading List - 13:34 - Next Week

Read aloud with edge-tts (en-US-AndrewNeural)

---

Story of the Week

First Documented Case: US Government Forces Emergency Shutdown of a Commercial AI Model

Three days. That’s all Anthropic’s most powerful model got.

On June 9, 2026, Anthropic launched Fable 5 and Mythos 5 as the most capable models in company history – setting new records in coding, reasoning, and multimodality [1]. Then came June 12, 5:21 PM Eastern Time: a directive from the US Department of Commerce forced Anthropic to immediately and globally block both models [2].

The government’s official justification: a potential jailbreak bypass classified as a national security risk [2]. From the agency’s perspective, the risk calculus was clear: an uncontrolled jailbreak in an already widely deployed frontier model could be exploited by state actors – and since selective deactivation was technically impossible, only one option remained. Whether this assessment proportionally justifies the measures taken remains publicly unanswered. The technical problem is structural: Anthropic cannot selectively disable the models for specific user groups – a complete global shutdown was the only available option [2]. Not only US users are affected, but all international customers and foreign employees as well. Mythos-class customer data is retained for 30 days [2].

Anthropic complied with the directive while clearly objecting to it:

“We disagree that finding a narrow potential jailbreak should be grounds for recalling a commercial model distributed to hundreds of millions of people.” [2]

— Anthropic, official statement on the government directive

Internally, Amazon CEO Andy Jassy reportedly raised safety concerns even before the regulatory directive, as TechCrunch reported on June 13 [11]. Amazon is one of Anthropic’s strategic cloud partners via AWS – making this dynamic particularly significant.

Why This Affects Everyone: This is the first documented case of a US government agency stopping an already commercially deployed frontier model via an emergency directive. This sets precedents in three dimensions: for AI export control as a regulatory instrument, for the architectural question of whether selective deactivation will become mandatory in the future, and for the liability of model providers vis-à-vis government authorities. Anthropic had reportedly pursued a “multilayered defense strategy” [2] – the fact that this wasn’t enough is the real signal.

Open Questions: How long will the block remain in effect? Can Anthropic patch the jailbreak and apply for reinstatement? And: will other governments – above all the EU – use this lever as well?

Bottom Line: If even Anthropic – the company that considers safety its core mission – cannot prevent a government emergency shutdown, AI export control has become a reality.

---

More Top Stories

Apple Turns Siri into a Gemini App – iOS 27 for Everyone from iPhone 11

Apple rebuilt Siri from the ground up at WWDC 2026 – with Google Gemini as the AI engine under the hood [3]. The partnership with a direct rival in the AI race is an unusual strategic move: Apple gains Gemini’s reasoning capabilities, while Google gets access to over one billion active iOS devices. At the same time, iOS 27, iPadOS 27, and macOS 27 “Golden Gate” were announced [3].

iOS 27 is the most inclusive iOS version ever: compatible from iPhone 11, with 70% faster photo display and 80% faster AirDrop transfers [3]. Tim Cook said farewell at WWDC – his successor as CEO will be John Ternus, planned for September 2026 [3]. However, the Gemini integration remains blocked in the EU and China for now – regulation is stalling the rollout in two of the most important markets.

Craig Federighi put Apple’s stance succinctly: “We believe privacy in AI is non-negotiable.” [3] – a statement that raises interesting questions in the context of a Google model at the core of the platform. Apple has not yet communicated details on how data processing works between iOS and Google’s Gemini infrastructure.

So What? For developers, Gemini-in-Siri means Apple Intelligence finally becomes usable. For Google: the world’s largest distribution channel. For everyone else: which manufacturer will be next to rely on a third-party LLM instead of training their own?

---

OpenAI: $852B Valuation, $85B Loss, One Attorney General Subpoena

OpenAI has confidentially filed an S-1 prospectus with the US Securities and Exchange Commission – one week after rival Anthropic, with Goldman Sachs, Morgan Stanley, and JPMorgan as lead bankers, and a planned IPO in September 2026 [4]. The leaked numbers are remarkably contradictory: approximately 900 million weekly active users, a valuation of $852B (March 2026) and $880B on the secondary market (April 2026) [4] – alongside a projected annual loss of $85B for 2028 with $122B in compute costs alone [4].

At the same time, US state attorneys general led by New York served OpenAI with a subpoena just before the IPO [7]. Requested: documents on advertising practices, user engagement, health data, and minors’ protection – supplementing an ongoing lawsuit from Florida [7].

So What? If the world’s largest AI company at ~$900B valuation and 900 million users will still lose $85B in 2028, the key IPO question isn’t “growth?” but “when does the model flip to profitability?” – and the answer is buried in an S-1 the public isn’t allowed to see yet.

---

Jeff Bezos’ Prometheus: $12B for the AI Engineer of the Physical World

Out of stealth mode and straight into the headlines: Prometheus, founded by Jeff Bezos and Vik Bajaj (ex-Verily/Google X), has gone public with a new $12B funding round and a $41B valuation [5]. The goal is an “Artificial General Engineer” – AI software that fully automates the design and manufacturing of physical systems like jet engines and pharmaceutical compounds. Offices in San Francisco, London, and Zurich [5].

The ratio is staggering: just 150 employees at a $41B valuation [5]. Investors include JPMorgan, BlackRock, and Goldman Sachs. Bezos himself predicted at the announcement “labor shortages rather than job losses”: “Significant productivity in the economy is going to raise the standard of living.” [5] Critics point out that productivity-enhancing technologies have historically first eliminated existing jobs before new fields emerge – a transition dynamic Bezos’ statement glosses over. The startup builds on an earlier stealth round of $6.2B (late 2025) – total capital raised amounts to approximately $18B [5].

So What? When biotech and aerospace veterans from Verily and Google X join forces with Bezos capital to bet on physical AI, this is no longer a research project – it’s the next industrial bet, with $41B in chips on the table.

---

Quick Hits

In brief:

• Mistral: Europe’s leading AI startup plans a €3B round at a €20B valuation – nearly doubling from the last Series C; sovereign wealth funds and global VC syndicates as investors (Bloomberg-confirmed) [6].
• Meta/Manus: Meta is unwinding its $2B deal with Chinese AI agent startup Manus after Beijing imposed conditions Meta found unacceptable – a textbook case of geopolitical risks in US-China AI investments [8].
• AI Music Flood: 44% of all songs uploaded daily to Deezer are AI-generated – around 75,000 tracks per day, of which approximately 85% of streams are classified as fraudulent (fake artists, royalty manipulation) and demonetized according to Deezer [9].

---

Tool of the Week

Deezer AI Music Detector – Automatically identifies AI-generated songs in your streaming playlists

Deezer has launched a free online tool that detects AI-generated music across platforms in playlists – compatible with Spotify, Apple Music, SoundCloud, and YouTube Music [9]. The tool supports 27 languages and is designed to simplify licensing and labeling processes for labels, platforms, and rights holders. Deezer itself already removes AI tracks from recommendations and editorial playlists and demonetizes fraudulent streams [9].

Particularly useful for music producers, playlist curators, and anyone who wants to know how much of their favorite playlist still comes from human hands. What sets Deezer apart from alternatives: it’s the only publicly available, cross-platform tool of its kind. CEO Alexis Lanternier: “No other company has followed our lead yet, so we decided to give everyone control over their playlists.” [9]

Deezer AI Detector – TechCrunch Article

---

Fail of the Week

“We expect our employees to use AI responsibly – including human review.”

KPMG had to withdraw its report “Redefining excellence in the age of agentic AI” this week after AI detection tool GPTZero identified inaccuracies [10]. Several organizations named in the report – UBS, NHS (UK), Swiss Federal Railways, and Transport for London – confirmed in writing that statements attributed to them were false or misleading [10]. The report apparently contained AI-generated statistics and citations that were adopted without human verification. Not an isolated incident: EY also withdrew a report in May 2026 due to AI-hallucinated footnotes [10].

What makes this fail particularly notable: KPMG sells AI governance consulting to its clients – and failed to follow its own guidelines.

Root Cause: LLM output was directly incorporated into the final report as research findings, without anyone contacting the cited organizations or verifying the stated figures against primary sources. Classic copy-paste from model output without factual verification.

What We Learn: Every figure, every quote, every company name in an AI-generated document requires manual verification against primary sources – especially when the document bears the name of a Big Four firm.

---

Number of the Week

$85,000,000,000

OpenAI’s projected annual loss for 2028 – despite approximately 900 million weekly users and a valuation of $852B [4]. In the same year, the company plans to spend ~$122B on compute alone – more than the GDP of Slovakia [4]. These figures are contained in the confidentially filed S-1 prospectus for the planned IPO in September 2026.

For comparison: Amazon needed nine years to reach its first profitability. For OpenAI, the critical question hinges on whether compute costs fall faster than user growth drives losses – and whether investors are willing to accept that for a $852B valuation.

---

Reading List

For the weekend:

1. Anthropic: Fable 5 & Mythos 5 – Launch Announcement – What Fable 5 and Mythos 5 can do technically – before the government pulled the plug. Essential reading to understand what was blocked here and why the jailbreak allegation is so controversial. (8 min)
2. TechCrunch: OpenAI S-1 – What the Leaked Numbers Mean – The $85B loss projection for 2028 in context: why this raises a fundamental profitability question for AI investors, not a short-term growth story. (10 min)
3. TechCrunch: Prometheus – Bezos’ Bet on the Physical AI Engineer – Why a startup with 150 employees at a $41B valuation wants to redefine engineering – and what that means concretely for manufacturing, pharma, and aerospace. (7 min)

---

Next Week

What’s coming:

• Anthropic/Fable 5: The critical question: can Anthropic patch the jailbreak within the next 7–10 days and apply to the Department of Commerce for reinstatement? A formal justification from the agency is expected mid-next week.
• OpenAI IPO: Following the attorney general subpoena, statements from lead bankers (Goldman Sachs, Morgan Stanley) are expected – and whether the September timeline still holds.
• Apple WWDC Developer Sessions: Technical sessions on Gemini-in-Siri and the Apple Intelligence APIs continue; first SDK documentation for external developers is expected by end of week.

---

🤖 Behind This Newsletter

Generated in: ~35 minutes Sources scanned: 17 articles from 6 feeds Stories found: 17 → 7 selected Validation: 4 agents, 11 issues → 11 corrected Model: Claude Sonnet 4.6 + Haiku (Validation) Images: Pollinations.ai (5 generated)

Full Metrics

Phase	Metric	Value
Source collection	RSS feeds	6
Source collection	WebSearch queries	4
Selection	Stories presented	17
Selection	Stories selected	7
Draft	Words	~1,500
Draft	Sources cited	11
Validation	Fact-check issues	3
Validation	Balance issues	4
Validation	Quality issues	4
Validation	Legal issues	0

---

This newsletter was researched and written AI-assisted. Images generated with Pollinations.ai.