The US government ordered Anthropic to shut down access to its most powerful AI models. Anthropic complied — but 200 organizations in a cybersecurity program still have access to the raw, unguardrailed version. The Permission Layer just got very real.
The Numbers
3 days
How long Fable 5 was public before the government pulled the plug
200
Organizations that still have access to Mythos Preview
23,019
Vulnerabilities Mythos flagged in open-source projects
On June 9, Anthropic launched Fable 5 — the guardrailed commercial version of its most powerful AI model, Mythos. Three days later, on June 12, the US government sent Anthropic a letter ordering it to block both Fable 5 and Mythos 5 from use by any foreign national, whether inside or outside the United States, citing national security concerns.
The government believed it had “become aware” of a method to jailbreak Fable 5 — bypassing the safety guardrails that were supposed to prevent the model from being used for offensive cybersecurity or bioweapon research. Anthropic complied immediately, shutting off access to both models for all customers to ensure no foreign nationals retained access.
But here’s the twist: Bloomberg now reports that approximately 200 organizations in Project Glasswing — Anthropic’s cybersecurity program — still have access to the raw Mythos Preview. Not Fable 5 with guardrails. Mythos itself.
The paradox: The US government shut down the guardrailed version (Fable 5) because it could be jailbroken. But 200 organizations still have access to the unguardrailed version (Mythos) — because they were cleared to use it for cyber defense.
Project Glasswing — The Cyber Defense Program
Project Glasswing launched in April 2026 with 50 initial partner organizations cleared by Anthropic to use Mythos Preview — the raw, unguardrailed model — specifically to scan codebases for security vulnerabilities. In June, Anthropic expanded the program to 200 organizations, including infrastructure operators in power, water, healthcare, and telecommunications.
The results have been extraordinary. Early partners collectively identified more than 10,000 high- or critical-severity security flaws. Anthropic used Mythos to scan more than 1,000 open-source projects, flagging 23,019 potential vulnerabilities — 6,202 of them estimated as high or critical.
Among confirmed participants: Cisco Systems and Dragos Inc. (whose CTO Jon Lavender confirmed access). These are companies trusted with critical infrastructure security — and they still have the most powerful AI model in the world at their disposal.
US government orders shutdown — jailbreak discovered. Fable 5 + Mythos 5 cut off for all customers.
June 2
Glasswing expanded to 200 orgs (before shutdown)
June 19
Bloomberg reports: 200 Glasswing orgs still have Mythos Preview access
The Structural Read
This story is not about a policy loophole. It is about the emergence of a two-tier AI access regime — and Anthropic is at the center of both tiers.
Tier 1: Public access (shut down). Fable 5 was available to everyone for three days. Then the government pulled it because the guardrails could be bypassed. The message: the public cannot be trusted with frontier models, even guardrailed ones.
Tier 2: Cleared access (still running). 200 organizations have the raw model — no guardrails — because they were vetted by Anthropic and cleared for defensive cybersecurity use. The message: access to frontier AI is becoming a security clearance, not a product feature.
The Emerging Reality
Frontier AI is becoming a controlled substance. Anthropic is becoming its licensed distributor.
Three Implications
1. ANTHROPIC’S SAFETY REPUTATION IS NOW A MOAT
The government shut down Fable 5 — but let Glasswing continue. Why? Because Anthropic spent years building trust with the national security establishment. The safety-first doctrine that critics called a marketing ploy just became the reason 200 organizations get to keep the most powerful AI model on Earth. Safety is not a constraint on Anthropic’s business. It is the business.
2. THE PERMISSION LAYER IS NOW A GOVERNMENT FUNCTION
We theorized the Permission Layer — the regulatory and credibility gate that decides which AI systems get deployed. This week it became real. The US government is now deciding who gets access to frontier models and who doesn’t. This is export control for intelligence, not hardware.
If the standard is that frontier models must be cleared for foreign access, OpenAI’s GPT-5 and Meta’s open-source Llama face the same risk. Anthropic argues this exact point — that the directive “would halt all new frontier model deployments across the AI industry.” Anthropic may have been first, but it won’t be alone.
The Bigger Picture
This is the same week John Jumper — Nobel laureate, AlphaFold creator — left Google DeepMind for Anthropic. The same week Trump said at the G7 that Anthropic talks are “going fine.” The same company that just had its model restricted by the government is simultaneously hiring the most prestigious scientist in AI and negotiating with the White House.
Anthropic is becoming something new in the AI industry: a company that is both constrained by the government and empowered by it. Fable 5 is shut down for the public — but Mythos runs freely for the defense establishment. That is not a contradiction. That is the businessmodel.
Permission Layer Doctrine
Access Is the New Moat
In the Permission Layer framework, the question is not “who has the best model” but “who is allowed to deploy it.” This week, that question was answered by a government letter. Anthropic’s safety investments didn’t prevent the shutdown — they ensured that Anthropic was the company trusted enough to keep the program running anyway. The Permission Layer is not a theory anymore. It is a fact.
The Bottom Line
Frontier AI models are becoming controlled assets. Access is splitting into public (guardrailed, revocable) and cleared (raw, persistent). Anthropic — through years of safety work, government relationship-building, and programs like Glasswing — has positioned itself as the licensed distributor for the cleared tier. That is not a weakness. It is the most defensible position in the AI industry. The model that gets shut down for the public but keeps running for the Pentagon is not losing. It is winning the game that matters.
Business Engineer Deep Dive
The AI Supercycle — Nine Layers, Three Clocks
The Permission Layer sits above the model layer in the AI Supercycle. This week proved it: government access control, not model capability, determines who captures value. The complete map of all nine layers.
Gennaro is the creator of FourWeekMBA, which reached about four million business people, comprising C-level executives, investors, analysts, product managers, and aspiring digital entrepreneurs in 2022 alone | He is also Director of Sales for a high-tech scaleup in the AI Industry | In 2012, Gennaro earned an International MBA with emphasis on Corporate Finance and Business Strategy.
Scroll to Top
Discover more from FourWeekMBA
Subscribe now to keep reading and get access to the full archive.
