Last night at 5:21pm ET, Anthropic received an export control directive from the US government and, within hours, disabled Fable 5 and Mythos 5 for every customer. Not throttled. Not geofenced. Off. The stated trigger is a national security concern tied to a reported method for bypassing Fable's safeguards. The order names foreign nationals, inside or outside the US, including Anthropic's own foreign national employees, and the only way to comply was to pull both models for everyone.
If you have been reading this newsletter, you know why this lands in a product security inbox and not just a policy one.
We have spent issues on the difference between a dedicated cyber model and a general purpose one. Mythos is the dedicated one. It ships with its own skills and harness, the way Claude Code does, and it does vulnerability discovery work out of the box. Fable is general purpose. It can do the same class of work, but only when you wrap it in a discovery pipeline, the scanner, process, runtime, exploit, and validation agents we walked through in Issue 2. Most people assumed only the cyber model would ever draw regulatory fire. The directive took both. The general purpose model got pulled right alongside the specialist one, because the capability lives in the reasoning, not the label on the product page.
That is the part worth sitting with. The thing that got Fable recalled was, per Anthropic's own statement, a narrow jailbreak that amounts to asking the model to read a codebase and fix the flaws in it. Anthropic says it validated that the same capability is available from other models, including OpenAI's GPT-5.5, and is used every day by defenders. Whether or not you agree with the government's read, the line between "general purpose assistant" and "vulnerability discovery tool" was thin enough that one directive erased it.
For a year we have run testing pipelines on frontier models, and we have been blocked by guardrails, forced to backtrack onto older models mid-run, and made to rebuild around capability that quietly moved or disappeared. That was our problem to absorb. It rarely made the news.
What changed last night is the size of the affected group. A model that hundreds of millions of people touch went dark on a few hours' notice, for a reason most of those users will never see the evidence for. Every business that had wired Fable or Mythos into an operation found out, in real time, what it feels like when access to a frontier model is a privilege the government can revoke between one afternoon and the next.
This is the early glimpse into the future we keep gesturing at. The dependency is real, the off switch is real, and it does not belong to you.
Read the wording of the directive closely. It is built around nationality. Foreign nationals are the restricted class, US persons are not the stated target. So if Anthropic wants to restore access to anyone fast, the path of least resistance is US persons first.
I don't know the mechanism, and I'd bet Anthropic doesn't fully either yet. But the obvious shape is identity verification. If you want Fable back, you prove you are a US person, probably by submitting an ID. That is a real possibility, and it carries consequences that nobody has priced in.
It would mean any non-US-citizen working on Fable is benched until the order lifts or a carve-out appears. Teams that are not all US persons do not get a clean half-restoration. They get a compliance problem.
It would also mean account sharing becomes a fast way to lose an account. If a verified US person hands their access to a non-citizen colleague or friend, that is the exact thing the order exists to prevent, and I'd expect detection and bans to move quickly. The 30-day data retention Anthropic already requires on Mythos-class models gives them the telemetry to notice.
I could be wrong on the mechanism. I am fairly confident on the direction. When access is rationed under an export rule, it gets rationed by who you are, and that has not been a variable most product security teams modeled into their tooling.
Treat frontier model access the way you already treat a critical single-source supplier, because that is what it is now. If a directive, a jailbreak report, or a quiet capability change can take your discovery pipeline offline on a few hours' notice, that is a continuity risk, and it belongs in your risk register next to the chip shortage and the sole-source sensor.
Keep your harness model-portable. We have said this before and last night made the case for us. The agents, tools, handoffs, and validators are yours. The reasoning engine in the middle should be swappable, because the one you are using today can be legally unavailable tomorrow.
And do not let the convenience of a model that ships with its own cyber skills lull you into building hard dependencies on one vendor's continued willingness, or legal ability, to serve you. The capability is not going away. The specific door you walked through to reach it might.
We will get into how to actually build a model-portable discovery pipeline, the Fable case rather than the Mythos one, in a future issue. Last night just made it the urgent one.