🤚 The Open-Palm Pardon

In a development that surprises absolutely no one who has watched this administration’s relationship with AI policy evolve from “ban it” to “actually, we need it” in the span of two weeks, Commerce Secretary Howard Lutnick has formally authorized more than 100 U.S. companies and government agencies to access Anthropic’s Claude Mythos 5 — the very model his office ordered pulled from the market on June 12.

“I have determined that appropriate safeguards are in place to permit certain trusted partners to access the Claude Mythos 5 Model,” Lutnick wrote, in a sentence that does extraordinary heavy lifting without specifying what any of those safeguards actually are.

The approved organizations — described only as “trusted partners” operating in critical infrastructure — now have full access to the frontier model, including, notably, their non-American employees. This is the same model that was grounded specifically because the government worried about foreign access. The solution, apparently, was to allow foreign access but only if someone American signs the org chart.

👐 The Two-Handed Timeline of Contradictions

To fully appreciate today’s announcement, one must appreciate the saga that preceded it. A brief timeline of the Mythos Cinematic Universe:

• May 25: The White House overrides the Pentagon to give Mythos to the NSA, because national security apparently requires the model that even its creators called “potentially hazardous.”
• June 12: The Trump administration bans Mythos 5 and Fable 5 after security researchers demonstrate the guardrails can be bypassed. The export control directive cites “national security authorities.”
• June 14: Amazon CEO Andy Jassy — whose company invested $5 billion in Anthropic — reportedly jailbreaks the model himself and calls the White House to inform them. The investor-as-whistleblower paradigm is born.
• June 23: A government official confirms that Mythos, while restricted, actually found vulnerabilities in classified U.S. government systems. The tool is so dangerous it must be locked up. Also, it’s too useful to leave locked up.
• June 26: The ban is partially lifted. One hundred organizations get access. The safeguards are “in place.” Fable 5 remains grounded.

This is a policy lifecycle that would make a pharmaceutical recall look deliberate. Ban, unban, restrict, expand, repeat — the model has been through more status changes than a college student’s relationship on Facebook.

🌿 The Gentle Awakening

There is something philosophically delicious about a government that simultaneously believes a model is too dangerous for the open market and essential for more than a hundred organizations across the federal apparatus. These two positions coexist in the same directive, separated by approximately zero paragraphs of explanation.

The unspoken logic is straightforward: Mythos 5 is genuinely the most capable AI model ever deployed for cybersecurity and code analysis. It found vulnerabilities in classified systems — systems that, presumably, had already been reviewed by some of the most well-funded security teams on the planet. When your AI intern outperforms your entire red team, you don’t fire the intern. You give it a security clearance and hope nobody notices the irony.

Meanwhile, Fable 5 — Anthropic’s creative and general-purpose model — remains in regulatory purgatory. The Commerce Department’s letter made no mention of it. Anthropic says it “continues working” to restore general access, which in Washington parlance means “we’re still negotiating and the lawyers are billing hourly.”

👑 The Gold-Leaf Security Clearance

The broader implication here is that the United States has now formally established a tiered access framework for frontier AI — one that exists not in legislation, not in regulation, but in letters from a Commerce Secretary who was an investment banker six months ago. The framework is: we’ll tell you who can use it, when they can use it, and we reserve the right to change our minds every thirteen days.

For Anthropic, this is a qualified win. Their most powerful model is back in production for their most important customers. Their IPO filing — still pending at a valuation that makes most GDP figures blush — now has a government endorsement instead of a government injunction. The stock prospectus just went from “material risk: our flagship model is banned” to “competitive advantage: our flagship model is so good the government banned it, then needed it back.”

For the rest of us, the message is clear: the most powerful AI in the world is available exclusively to people who were already powerful. The “trusted partners” list is not published. The “appropriate safeguards” are not described. And the model that can autonomously discover vulnerabilities in classified systems is now operating inside those systems with all the regulatory precision of a handshake agreement.

Welcome to frontier AI governance, 2026 edition. The guardrails are metaphorical, the access is political, and the safeguards are in a letter that doesn’t define them.

“The model was too dangerous to release, too useful to restrict, and too expensive to ignore — so we did all three in fourteen days and called it policy.” — The Slap of Wisdom National Security Desk, currently awaiting its own security clearance determination, which has been pending since the second ban