A Friday Letter From Washington Took Anthropic's Two Best Models Offline

San Francisco | Monday, June 15, 2026

A single Commerce Department letter took Anthropic's two strongest models offline worldwide on Friday, the first time Washington has switched off a generally available American AI model. The trigger was a jailbreak Anthropic calls narrow, one that rival models answer too.

For enterprise buyers, continuity just joined capability and price on the shortlist. Our op-ed argues that the power to pull a model used by millions belongs with a standing expert board rather than a Cabinet fighting Anthropic in court.

Meanwhile Meta is generating its own crisis. Zuckerberg concedes mistakes while the engineers he drafted into a three-month-old AI unit call the work a gulag, and the company caps the token spending it spent a year pushing.

Stay curious,

Marcus Schuler

Anthropic Pulls Fable 5 and Mythos 5 Worldwide After Commerce Department Order

Anthropic disabled Claude Fable 5 and Mythos 5 for every customer worldwide on Friday, after a Commerce Department order barred all foreign nationals from the two models. Washington has never pulled a generally available U.S. AI model before.

The export-control directive required a license for any transfer and carried civil penalties, Axios reported, and because Anthropic cannot screen users by nationality in real time, it shut both models off three days after launching Fable 5 as its most capable yet. Anthropic calls the cited jailbreak narrow and says OpenAI's GPT-5.5 produces the same output. The Wall Street Journal traced the trigger to Amazon, whose chief executive told officials his researchers had pushed Fable into cyberattack-relevant territory.

Why This Matters:

• Enterprises running critical workflows on a single closed-API model now face a regulatory single point of failure, not only an uptime or pricing risk.
• Anthropic filed confidentially for an IPO this month at a $965 billion valuation, and a prospectus would force this shutdown risk into investor line items.

Anthropic Pulls Fable 5 and Mythos 5 on US Government Order

A single Commerce Department letter took Anthropic's two most powerful models offline worldwide, the first time Washington has pulled a generally available AI model off the market. For enterprise buyers, continuity just became a question that ranks beside capability and price.

Implicator.ai

The One Number

💰 Fresh Funding

Take AI's Kill Switch Away From the Politicians

Our opinion piece makes the harder case on the Anthropic shutdown: the government can take a tool away from the world, but it should at least have to show its work.

Security researcher Katie Moussouris called the order "a complete overreaction" after reading the report officials acted on. The trigger was a researcher asking Fable to read a codebase and patch flaws, the daily work of network defenders, output she told the Wall Street Journal would help defenders more than attackers.

Washington tried this in the 1990s, classifying strong encryption as a munition and investigating PGP's author until a federal court ruled source code is protected speech. The op-ed's answer is a standing board of AI researchers, security scientists, and constitutional lawyers, walled off from the administration and ruling in the open. A board like that could still pull a dangerous model, with its reasons on the public record.

Anthropic Mythos Ban Shows AI Needs an Expert Board

The U.S. government made Anthropic pull its two most capable AI models over a narrow jailbreak that defenders use daily. Who can switch off a model used by millions, and on what evidence, should not sit inside a Cabinet that is also the company's antagonist.

Implicator.ai

AI Image of the Day

Prompt: A model-looking woman of African-American descent, natural appearance, in the city, wearing Desu Clothing.

Zuckerberg Admits Mistakes as Meta's 6,500-Person AI Unit Nears Revolt

Mark Zuckerberg told staff Friday that Meta "made mistakes and will almost certainly make more" rebuilding the company around AI, Reuters reported.

Days earlier an employee hijacked a livestreamed all-hands to call a Meta AI executive "a piece of shit," and engineers inside the three-month-old, 6,500-person Applied AI unit, many reassigned without a choice, were calling themselves draftees and their work, generating coding puzzles to train models, a gulag.

The same week Meta moved to cap the employee token spending it had encouraged for months, after staff burned 60 trillion tokens in 30 days on a leaderboard called Claudeonomics. Zuckerberg offered assigned desks and no fresh layoffs this year, fixes aimed at the conditions, not at a strategy that still needs elite engineers writing puzzles by hand.

Zuckerberg Admits Mistakes as Meta's AI Unit Nears Revolt

An employee hijacked a livestreamed Meta meeting to insult a top AI executive, exposing a revolt inside its 6,500-person Applied AI unit. The same week, Meta moved to cap the token spending it spent months pushing. Zuckerberg admits mistakes, but the contradiction runs deeper.

Implicator.ai

🧰 AI Toolbox

How to Build a Personal ChatGPT That Runs Entirely on Your Own Machine With Remio

Remio is an on-device AI app that captures the documents, screenshots, web pages, and notes you encounter every day, builds a searchable personal knowledge base, and answers questions across all of it locally. Nothing leaves your device unless you opt in. Useful for anyone who wants the convenience of ChatGPT with their own data without sending it to an API. Free download for Mac and Windows.

Tutorial:

1. Download Remio from remio.ai for Mac or Windows
2. Pick a local model during setup based on your hardware (smaller models for laptops, larger for desktops with 32GB+ RAM)
3. Turn on capture sources: browser pages, screenshots, PDFs, Markdown notes, or a watched folder
4. Let Remio index in the background, the knowledge base grows automatically as you work
5. Open the chat and ask a question across everything you've captured: "What were the key points from the Stripe API doc I read last week?"
6. Use the timeline view to scrub through what you captured on any given day, with clickable thumbnails
7. Enable selective cloud sync if you want to access your knowledge base from another device, with end-to-end encryption

URL: https://www.remio.ai

What To Watch Next

💡 5-Minute Skill: Turn a Job Offer Into a Counter You Can Defend With Numbers

Monday, 9:05 a.m. The offer landed over the weekend and the recruiter wants an answer by Wednesday. Before you accept out of relief, find the levers worth negotiating.

Your raw input:

The prompt:

The output:

Counter base at $200,000, anchored to the $195K to $210K range; you are under market. Ask the equity strike price and cliff before valuing the RSUs. Request two remote days, and trade the signing bonus for base. Reply: "Could we align base nearer $200,000 and confirm equity details Tuesday?"

Why this works:

Most offers get accepted on relief, not analysis. This splits the package into levers, attaches a market number to each, and names the one to concede, so you bargain from evidence instead of nerve.

What to use:

Claude is best when you paste the full offer and your research. ChatGPT is fine for polishing the reply. Keep the phrase "the one lever to trade away," or the model asks for everything and you sound greedy.

📖 AI Alphabet

AI & Tech News

FBI Dismantles AI-Powered Chinese Phishing Service Spanning a Million URLs

The FBI, working with Google and Black Lotus Labs, took down Outsider Enterprise, a phishing-as-a-service operation tied to China, and seized about $100,000 in Tether. The service used AI to generate and scale malicious emails across more than a million URLs that impersonated real brands to steal credentials and plant malware.

ByteDance Courts Chinese Chipmakers Iluvatar CoreX and Baidu for AI Inference GPUs

ByteDance is in talks to buy AI inference chips from Shanghai-based Iluvatar CoreX and is weighing a separate deal for Baidu's Kunlunxin silicon, sources told Reuters. The move would diversify its hardware supply as U.S. export limits push Chinese firms toward domestic chips.

Cybersecurity Leaders Press the White House to Reverse the Anthropic Model Ban

A coalition of CISOs and researchers from Adobe, Zoom, and Sophos urged the administration to lift the restrictions on Anthropic's Fable 5 and Mythos 5, arguing the ban hurts defenders more than attackers. Their June 15 letter says security teams rely on the models for threat detection, incident response, and vulnerability analysis.

AI 'Nudify' Tools Drive a New Wave of School Bullying

The spread of AI nudify tools is letting students fabricate explicit images of classmates, the Wall Street Journal reports, creating a new form of cyberbullying. Schools and parents are scrambling as photos pulled from social media get turned into realistic deepfake nudity, and calls for platform accountability are growing.

A Top Deepfake-Detection Expert Says He Can No Longer Tell What Is Real

UC Berkeley's Hany Farid, a leading digital forensics authority for two decades, says AI-generated media has outpaced his field, according to the New York Times. Farid describes being fooled by realistic AI forgeries himself, a turning point in the effort to catch synthetic content.

Trump Threatens 100% Wine Tariffs if France Keeps Its Digital Tax

In an interview with the New York Post, President Trump said he warned Emmanuel Macron to repeal France's 3% digital services tax on U.S. tech firms or face a 100% tariff on French champagne and wine. The threat revives a long-running fight over levies Washington views as aimed at American companies.

Britain Moves to Bar Under-16s From Social Media by 2027

Prime Minister Keir Starmer announced a ban on social media for children under 16, covering TikTok, Instagram, YouTube, Snapchat, Facebook, and X, with legislation targeted for 2027. Platforms would have to verify age and block minors from livestreaming under broader child-safety duties.

Tech Billionaires Coordinated in a Signal Chat to Kill California's Wealth Tax

Leaked messages show Sergey Brin, Marc Andreessen, Garry Tan, Chris Larsen, Mike Moritz, and Ron Conway organized opposition to a proposed 5% wealth tax in a private Signal group, the San Francisco Standard reports. The group weighed funding campaigns and shaping public messaging, an effort that helped defeat the measure at the polls.

Rylo Raises $85M to Scale AI Sign-Language and Speech Translation

Israeli startup Rylo, formerly Nagish, raised $85 million in a Series B led by General Catalyst and Canaan to expand real-time speech-to-sign and sign-to-speech translation for deaf and hard-of-hearing users. The funding backs accuracy work across more sign languages and a global commercial rollout.

Hypha Lands $50M Seed to Automate Private-Market Underwriting

Hypha, which uses AI to extract and structure data from private-market documents, raised $50 million in seed funding led by General Catalyst. The platform automates underwriting, portfolio monitoring, and asset management for institutional investors working in opaque private markets.

🚀 AI Profiles: The Companies Defining Tomorrow

Apoha wants to give AI a data layer for the physical world: empirical measurements of how molecules and materials actually behave, not predictions inferred from sequence and structure. The London and San Francisco company came out of stealth on June 3 with a $36 million Series A and a flagship readout, VIBE, that fingerprints a sample in minutes. 🔬

Founders
Founded in 2021 by Shamit Shrivastava (CEO) and Anshika Srivastava (COO), a husband-and-wife team working the seam between science and finance. Shrivastava is a biophysicist with a Boston University PhD and an Oxford postdoc; his research on how matter behaves at liquid boundaries made Scientific American's 2018 list of discoveries that could change everything. Srivastava trained as an engineer at Dartmouth and ran engineering teams as an Executive Director at Goldman Sachs before co-founding the company.

Product
Apoha's VIBE assay takes a sample small enough to sit on a pinhead, suspends it in liquid, applies a controlled series of stresses, and records the wave patterns the molecule throws off in response. A single readout produces more than 1,000 empirically measured descriptors of behavior in minutes, where conventional lab instruments measure one property at a time. The company calls this Liquid State Intelligence and frames it as a third molecular data class alongside sequence and structure, fuel for AI that has to reason about real-world matter. It holds more than 60 patents across hardware, software, data, and models.

Competition
On measurement, Apoha runs against contract biophysics providers like Malvern Panalytical and Charles River Laboratories, plus scientific data platforms such as Dotmatics. Structure-prediction labs in the AlphaFold lineage forecast a molecule's shape; Apoha sells measured behavior under real conditions instead. Its wedge is owning a data type nobody else generates at scale, and it already runs pilots with Boehringer Ingelheim, Ethris, and food and materials customers.

Financing 💰
$36 million Series A led by Singular, with Tim Draper's Draper Associates joining alongside existing backers Redalpine, Seedcamp, Wilbe, and Nucleus, plus grant support from Innovate UK. Apoha employs about 29 people across London and San Francisco and reports customer results including better than 90% precision on antibody screening for Boehringer Ingelheim and lipid nanoparticle optimization for Ethris. Sources: Tech.eu, June 3, 2026; Tech Funding News, June 3, 2026.

Future ⭐⭐⭐
Materials and molecular discovery is the kind of market where measured data beats clever modeling, because the physical world refuses to be guessed. Apoha wins if Liquid State Intelligence becomes a standard input that pharma, food, and materials teams pay to generate rather than a niche assay. The risk is the usual deeptech tax: long sales cycles, instrument deployment, and the patience to prove a new descriptor set predicts outcomes the incumbents miss. 💧

🤨 Yeah, But...

Washington ordered Anthropic on Friday to cut off Claude Fable 5 and Mythos 5 to every foreign national on earth, including the company's own foreign-born engineers, three days after it launched Fable as the most capable model it had ever shipped. Security researcher Katie Moussouris, who read the report the government acted on, told the Wall Street Journal it was "a complete overreaction"; the triggering finding was a prompt asking the model to read a codebase and patch its flaws.

Sources: Axios, June 12, 2026 | Anthropic, June 12, 2026

Our take:

There is a long American tradition here.

In the 1990s Washington decided encryption was a munition and went after the man who posted PGP to the internet, right until a court noted the math was already everywhere and let it go. The 2026 update is that the munition now ships with a refund window.

A product used by hundreds of millions got reclassified as controlled technology over a single Friday evening, and the same capability is sitting in OpenAI's GPT-5.5, which nobody has asked to leave the building.

Anthropic says it is a misunderstanding and is working to restore access soon. Pre-IPO companies always say soon.

Marcus Schuler

San Francisco

Editor-in-Chief and founder of Implicator.ai. Former ARD correspondent and senior broadcast journalist with 10+ years covering tech. Writes daily briefings on policy and market developments. Based in San Francisco. E-mail: editor@implicator.ai