|
Last Friday at 5:21pm ET, Commerce Secretary Howard Lutnick sent a letter to Anthropic CEO Dario Amodei. By Saturday morning, the company's two most powerful models, Fable 5 and Mythos 5, were offline everywhere in the world. Five days from launch to global shutdown. That's the headline. The actual story is messier, more interesting, and a lot more relevant to how you should plan your AI roadmap for the next twelve months. Let me start by clearing up what actually happened, because the early coverage got it wrong. What actually happened (in 90 seconds) Anthropic launched Fable 5 and its more capable sibling Mythos 5 on June 9. Both were billed as the most powerful and most safety-tested models the company has ever shipped. Thursday night June 11, Amazon CEO Andy Jassy called Treasury Secretary Scott Bessent. Amazon's security team had been red-teaming Fable 5 and found that if you handed the model a codebase and asked it to identify vulnerabilities, it would do the work. Fable 5 functioned as a competent automated security researcher. By Friday afternoon, Commerce had its letter ready. The order: no foreign nationals can access Fable 5 or Mythos 5. Not abroad. Not inside the United States. Not even foreign-born Anthropic employees inside Anthropic's own offices. Anthropic had no way to verify the citizenship of hundreds of millions of users at scale, so it did the only thing it could do. It shut both models down globally. However, at it's core, the reaction is over something (the "jailbreak") that is, itself, contested. Anthropic published a response noting the same capability is available in OpenAI's GPT-5.5 without any bypass. Katie Moussouris, who runs Luta Security and has done red-team work for Microsoft and the Pentagon, told the Wall Street Journal: "It's not a jailbreak. This is a complete overreaction." Why this is a bigger deal than a single model recall Three layers. Layer one: precedent. A cabinet secretary called a US company on a Thursday night, and by Friday afternoon two of its products were globally unavailable. No public comment period. No court order. No disclosure of what the underlying vulnerability actually was. That is now an executable playbook. Layer two: the retaliation question. In February, the Department of Defense designated Anthropic a "supply chain risk," the first time that label was applied to a US company. Anthropic sued. In March, Judge Rita Lin granted an injunction and used the phrase "classic First Amendment retaliation." Her finding: the designation appeared "driven by a desire to make an example of Anthropic for its public stance" on autonomous weapons and mass surveillance. The June export order may or may not connect to that earlier conflict. Reporters are noticing the pattern. Layer three: trust, in two directions. Anthropic has spent two years building a brand on safety and transparency. Last week the company was also caught silently routing queries in biology and cybersecurity to a weaker fallback model without telling users. Researchers called it "secret sabotage." Anthropic apologized: "We made the wrong tradeoff." So the trust hit is not only against the government. It is also against the lab that produced the model. Three scenarios for the next ninety days I'm tracking three. You'll hear arguments for others. These are the ones I think small business owners, CEOs, and investors should plan around. Scenario one: The Splinter. Fable 5 returns as "Fable 5 US," with verified-citizen access only. Other labs preemptively build the same architecture. The US AI market quietly bifurcates. There is a domestic-only frontier tier, and a global commercial tier, and your team starts routing workloads accordingly. This may be operationally difficult to build and enforce. What this means for you: vendor selection becomes a question about who your end users are. If you serve a global customer base, you cannot build on the frontier US tier. You'll route to OpenAI's international SKU, Google's, or one of the Chinese open-weight models for anything client-facing. Scenario two: The Capability Ceiling. Other labs read the room. OpenAI, Google, and xAI quietly delay their next frontier releases. "Fable-5-class" becomes a public threshold no one wants to cross. Frontier capability moves into private channels: defense contractors, national labs, enterprise-only tiers. What this means for you: the consumer-facing AI tools you use today are roughly what you'll have through the end of 2027. Plan your investments in AI infrastructure on that assumption. If your business case depends on "the next leap," reconsider. Scenario three: The Loyalty Tax. The retaliation framing becomes the dominant narrative. Other frontier labs notice the cost of refusing government use cases. OpenAI, Google, and xAI accept defense and surveillance contracts on more permissive terms. "Safety-first" stops being a brand asset and starts being a vendor risk flag. What this means for you: vendor due diligence now has a new question. "Is this AI vendor likely to be allowed to keep operating?" Board-level diversification conversations. A second-source clause in every AI contract. What to do this week Three things. One. If you have a single-vendor AI dependency in production, write down what a four-day cutoff would cost you. It is not a thought experiment anymore. Two. Ask your AI vendors, in writing, what their access continuity commitment is. Three. If you are an investor, the question on every AI portfolio company memo this quarter is the same. Which scenario are they exposed to, and what is their hedge? The Fable 5 story is not over. The next ninety days will tell us which version (or versions) of it we are living in. Alex |
As an AI Coach, Advisor, and Agent Builder, I help organizations and business leaders harness the power of artificial intelligence to boost productivity and streamline operations. I enable organizations to navigate the transformative landscape of AI, educating teams, identifying operational and strategic opportunities with AI and creating a framework for safe and transparent use of data in the organization.
Next week we're going to graduate from AI slop creator to real builder. I am going to show you how easy it is. We're going to build a To Do list tracker that really works *for you.* Why? Two reasons: One - Because I've downloaded, abandoned, and felt quietly guilty about more task apps than I can count. Todoist, Things, Notion, a legal pad, the Notes app, a whiteboard I can't see from my desk. None of them were wrong. They just weren't mine; each one was someone else's idea of how I should...
You've heard the complaints about AI slop. The generic LinkedIn posts. The Twitter threads that all sound the same. The flood of content that exists because it's now almost free to make. Everyone's pointing at the public version of this problem, the stuff strangers scroll past on their feeds. Nobody's talking about the slop we make for ourselves. Here's what I mean. The same thing that makes public AI slop so easy to produce (low effort, instant output, a little hit of delight every time) is...
Most of us ask AI to do the one thing it's worst at. We ask it to decide. "Which vendor should I pick?" "Is this a good hire?" "Should we raise prices?" And it answers instantly, confidently, in a clean paragraph. The problem is that the confident paragraph is often the shallow one. Here's how to fix that, and it takes about thirty extra seconds. First, the proof it matters. Researchers gave a top AI model a farmer's question: plant apples or grapefruit next year? The model saw that...