Washington Pulled Two AI Models Offline. Here Is What Happened, and What You Should Setup Now (1 Prompt)
A reported look at the Fable 5 and Mythos 5 suspension, the government’s case, Anthropic’s rebuttal, and a quick local setup guide in case actions like this become the norm.
You’re reading Excellent AI Prompts.
AI prompt chains, agent skills, and agentic workflows for professionals learning to use AI to earn more, think sharper, and live better.
Copy them. Use them. Build with them.
On the evening of Friday, June 12, the government pulled Anthropic’s most capable model.
According to Anthropic’s public statement, the company received an export control directive from the US government at 5:21pm Eastern that day. Axios first reported the source: Commerce Secretary Howard Lutnick had written to Anthropic chief executive Dario Amodei, placing the Mythos 5 and Fable 5 models under export controls covering any location outside the United States and all foreign persons inside it. Per the letter Axios described, a license would now be required to export, re-export, or even domestically transfer the models, with financial and civil penalties for failing to comply.
The directive named foreign nationals, including Anthropic’s own foreign national employees. Anthropic cannot tell in real time which of its users hold which passport, so the only way to comply was to shut both models off for the entire customer base. As the company put it, the net effect was that it had to abruptly disable Fable 5 and Mythos 5 for all customers, while access to every other Anthropic model continued normally.
The government’s case
An administration official told Axios that another company claimed it was able to jailbreak Mythos, which alarmed officials about national security risk. The administration had earlier tried to get Anthropic to delay releasing the new models and did not succeed, and the export control letter followed. The same official said the models needed to stay locked down until the government’s own national security systems were hardened, and suggested that could happen within the next few weeks.
Anthropic’s rebuttal
Anthropic’s account of the underlying problem is far smaller than the response. The company says the letter gave no specific national security detail. Its understanding is that the government believes someone found a way to bypass, or jailbreak, a Fable 5 safeguard meant to stop the model from helping identify software vulnerabilities.
The technique, as Anthropic describes it, was asking the model to read a specific codebase and fix the flaws in it. The company reviewed a demonstration and found it surfaced a small number of previously known, minor vulnerabilities, the kind other widely available models can find without any bypass. Anthropic says it validated that the level of capability on display is available elsewhere, naming OpenAI’s GPT-5.5, and is used every day by the security professionals who defend systems.
Anthropic also mentions the safeguards it built before launch. The company says Fable was red-teamed for thousands of hours with the US government, the UK AI Safety Institute, and outside groups, that no tester found a broad universal jailbreak, and that it adopted a defense-in-depth approach backed by a 30-day customer data retention policy, a costly choice with its enterprise customers, specifically so it could detect and shut down attacks quickly.
On the record, the company drew a hard line. “We disagree that the finding of a narrow potential jailbreak should be cause for recalling a commercial model deployed to hundreds of millions of people,” it wrote, adding that if the same standard were applied across the industry, it would essentially halt all new model deployments from every frontier provider. Anthropic called the episode a likely misunderstanding, said it was complying while it worked to restore access, and promised more detail within 24 hours.
Anthropic and Washington were already at odds
Anthropic already sits on a Pentagon blacklist that deems its models too risky for the government’s own use, and the company now also sits inside a Commerce licensing regime that deems them too risky for foreign use. Earlier in the month, the administration issued an executive order to test the most advanced models before deployment, a voluntary framework that pointedly avoided a licensing regime, a design White House AI adviser David Sacks pushed to prevent what he has called regulatory capture by the largest labs.
The rollout itself was uneven. The independent developer Simon Willison, who had been publishing hands-on notes about Fable 5 all week, flagged the statement as “nuts” and reported that he still had working access through Claude.ai and Claude Code at 9:01pm - 9:59pm Eastern, hours after the order landed. A directive can arrive faster than systems are able to enforce it.
What is verified, and what to make of it
A cabinet secretary’s letter took two deployed frontier models away from paying customers in a single evening. The government has not published its technical case. The company that built the models, with its own incentives, says the cited capability is ordinary and widely available. Both sides agree access was cut first and explained second.
You do not have to pick a side to learn the lesson. You don’t control any model, even if you are paying top dollar. You’re granted permission to use it, and that permission can be withdrawn by the provider, a court, or, as of Friday, a cabinet department, with no notice you can plan around. For a large company that risk is a hefty line item. For a one-person business it can be the entire business.
The local setup no one can take from you
I do think local, small, private models are a good strategic bet at least as a fallback.
Be honest about the limits first. A local model will not replace the frontier on the most complex work. It still trails the best cloud systems on the most demanding reasoning, the longest documents, and serious coding, and a large one needs real memory to run. For everyday business work, though, drafting, summarizing, reformatting, writing in your own voice, and answering questions over your own files, a good mid-size model in 2026 is already enough, especially once it can read your material.
The setup has three parts.
The computer. An Apple Silicon Mac is the simplest path, because its memory is shared with the graphics chip and models run well without extra parts. A machine with 32GB of memory runs a solid mid-size model, and 64GB or more runs the larger ones. A Windows or Linux desktop with a good graphics card works too.
The software. Ollama runs models from a single command, or LM Studio does the same through a normal app. On top of either, Open WebUI adds a chat window and local retrieval over your own files, so nothing leaves the device. Jan.ai is also an easy option for beginners.
The model. Qwen, Llama, and Mistral all ship open-weight versions that run locally, and a mid-size build is a good choice. The strongest specific version changes month to month, so tell AI your exact machine and it will check the current best fit before you install anything.
Set Up Local AI - A Copy-Paste Prompt and Resource List
Use this prompt on a cloud model
Fill in the bracketed parts, then paste the whole thing into any capable AI assistant. It will return a setup plan built for your exact machine.
You are helping me set up a local AI model on my own computer, so I have a
private model that keeps working even if a cloud provider goes down. Act as a
practical guide who has done this many times and explains things plainly.
Here is my situation:
- Operating system and version: [e.g. macOS 15, Windows 11, Ubuntu 24.04]
- Computer and chip: [e.g. MacBook Air M3, PC with NVIDIA RTX 4070]
- Memory (RAM), and GPU memory if you know it: [e.g. 16GB unified, or 32GB + 12GB VRAM]
- Free disk space: [e.g. 100GB]
- What I mainly want it for: [e.g. drafting client emails in my voice,
summarizing documents, bookkeeping notes, coding]
- How important privacy is: [e.g. some work is confidential client data that
must stay on my device]
- My comfort level: [e.g. I can copy-paste terminal commands but I am not
technical, or I prefer a clickable app]
Based on this, do the following:
1. Tell me whether my hardware is up to it, and what size of model it can run well.
2. Recommend ONE specific open-weight model to start with (name and exact
size or quantization), plus one lighter fallback if the first is too slow,
and explain the tradeoff in plain language.
3. Recommend ONE tool to run the model and ONE interface for chatting and using
my own files, and say why those fit my comfort level.
4. Give exact step-by-step install instructions for MY operating system, in
order, with the real commands or clicks. Assume I am starting from nothing.
5. Show me how to load my own files so the model can answer from them, and how
to keep that data fully on my device.
6. Give me a 5-minute test to confirm it works offline, including turning wifi off.
7. List the most common problems for my setup and how to fix each one.
Keep it concrete and sequential. Define any technical term the first time you
use it. If you need a detail about my computer that I did not give you, ask me
before guessing.
Local AI resources
Run a model on your own machine
Ollama (https://ollama.com): run open models from a single command. The simplest place to start.
LM Studio (https://lmstudio.ai): a clickable desktop app to download and run models, if you would rather avoid the terminal.
Unsloth (https://unsloth.ai): run open models locally and fine-tune one on your own data using far less memory. For when you want a model tuned to your own work.
Chat with the model and use your own files
Open WebUI (https://openwebui.com): a private chat interface that can read your own documents. Pairs with Ollama.
AnythingLLM (https://anythingllm.com): a desktop app that turns a local model into a private assistant over your own files.
Find and compare models
Hugging Face (https://huggingface.co): the main library of open-weight models to download, with each model’s size, license, and notes.
Arena (https://arena.ai): a leaderboard that ranks models by blind human preference, useful for seeing which open models are currently strongest before you pick one. Formerly known as LMArena.
A simple starting path
If you want one route without deciding everything yourself: install Ollama, browse Hugging Face or check Arena for a current strong mid-size open model in the Qwen, Llama, or Mistral families, pull it through Ollama, then add Open WebUI so you can load your own files. Reach for Unsloth later, once you want a model tuned to your own material.
Final Thoughts
You do not need to understand export controls to understand risk. When a single letter can turn your paid tools off between dinner and bedtime, it doesn’t invite much confidence in a stable future with AI cloud tools.
A local model you download to your own device is yours. No, they are not as capable as the frontier models, but they are yours.
You don’t need to abandon the cloud.
But I urge you to stop being dependent on it.
Test one critical task on a local model this week.
Until next time,
Lea
