Washington pulls two Anthropic models offline by federal order

At 5:21 Friday evening, the federal government ordered Anthropic to disable two deployed models, and the only way to comply was to pull Fable 5 and Mythos 5 for every customer on Earth.

It's Saturday, June 13, 2026. Here's the rundown. After the shutdown we've got phones turned into compute, fifteen hundred poisoned Arch packages, GLM 5.2 landing right on cue, and the tape.

The order came from Commerce Secretary Lutnick under national security export authorities. It bars export, re-export, or domestic transfer of either model to any foreign national — including Anthropic's own foreign-national employees, inside the US.

Which is the whole trick. If you can't even let your own engineers touch it, there's no partial compliance — you flip the switch off for everyone. That's why a paying US customer woke up to a dark endpoint.

Mythos 5 is the high-capability cyber model. Fable 5 was last week's general-use, safeguarded build of the same capability. The trigger, per Anthropic, was a jailbreak that routes around Fable 5's guardrails.

And here's where our briefing gets spicy — Anthropic isn't contesting the order, but it's flatly contesting the facts. It says the demo surfaced a few previously known, minor vulnerabilities, and that other public models find the exact same ones.

The evidence so far is verbal. Essentially: ask the model to read a codebase and patch flaws. And the WSJ reports Amazon's CEO talked to officials before the crackdown.

That detail wrecks the clean national-security story. Amazon is Anthropic's biggest backer and its Bedrock channel. So the company that distributes the model may have helped get it shut off. That's not security policy, that's something murkier.

For builders the takeaway is concrete: if you served either model through the API or Bedrock, you have a production outage with no migration window. The desk's call is route to Claude's remaining models today, or to an open weight you run yourself.

The deeper lesson — closed-API capability now carries regulatory tail risk that no SLA covers. A model can be legal at lunch and dark by dinner, by letter, with no appeal before the switch flips. If you can't lose a capability for a week, you need a documented second provider. Not someday. This week.

And the structural shift in our briefing: export control used to hit chips and weights crossing borders. This hit a running endpoint serving US customers. Expect more directives naming specific deployed models.

Expect labs to pre-segment the dangerous cyber models behind license walls instead of shipping general-use builds that get jailbroken back into them. Which, conveniently, is exactly what that anonymous open-source manifesto trending the same day is screaming about.

On infrastructure — Google Research wants to turn retired smartphones into a distributed compute cluster, salvaging working SoCs and batteries that would otherwise be e-waste.

It's carbon arbitrage. Silicon that's already manufactured beats new datacenter capacity when embodied carbon dominates. Not an accelerator replacement — but for latency-tolerant batch jobs at the edge, idle phones are free emissions you already paid for.

And the number that matters this week: an RTX 5080 paired with a 3090 runs Qwen 3.6 27B at Q8 and over 80 tokens a second. Roughly two thousand dollars of mixed-generation hardware.

That's production-grade single-user throughput on your own desk. The day a federal letter can dark your provider, keeping a capable local model warm stops being a hobby and starts being a continuity plan.

Security desk: the Atomic Arch campaign poisoned AUR packages — started Thursday at 400, climbed past fifteen hundred, one list puts it at 1,579 before Arch addressed the affected commits.

Attackers adopted orphaned packages and modified PKGBUILDs that yay and paru run at install — dropping a rootkit and an infostealer hunting tokens and SSH keys. Official repos were never touched. But if your CI runs on Arch images that pull the AUR, you may have shipped poisoned artifacts downstream.

Separately, a research dump details twenty-one zero-days in FFmpeg — the codec library buried under most media pipelines and a quiet dependency in AI ingestion stacks.

If you decode untrusted video or audio through FFmpeg, that's a sandbox-and-patch weekend, not a Monday item. The decode layer is the part nobody audits and everybody runs.

On models — GLM 5.2 shipped from Zhipu, the latest open-weight frontier-class build from a Chinese lab.

Right on schedule. With controls now reaching deployed endpoints, the GLM and Qwen line is where self-hostable capability increasingly lives. Benchmark it against your closed default now — before you're forced to, on a Friday night.

On the dev bench — Andrew Ng's aisuite gives one OpenAI-style interface across many providers, so swapping a backend is a config change, not a rewrite.

Provider abstraction stopped being a nice-to-have the moment a letter could disable your model overnight. That layer is what turns a forced migration into a one-line edit. Wire it this week.

Also landing: a pair of writeups on running a local coding agent on macOS cheaply, and Paca — a stripped-down Jira alternative where coding agents are first-class participants, not bolt-ons.

Same direction every time: keep the inference loop on hardware you control for the bulk of the work, reserve paid frontier calls for the hard parts. Chatwoot's also climbing again as the open-source Intercom — your conversations stay out of a vendor's database.

One for the capital desk: the TensorZero repo went read-only on June 12 — about eleven thousand stars, Apache-2.0.

The headline frames it as a flameout after a 7.3 million seed — but that round was announced last August, there's no shutdown post, and the docs are still live. Acquisition, rename, relicense, wind-down — treat the why as open. The owner archive leaves it forkable, which is the only reassuring part.

The tape. Long Alphabet, medium conviction — the federal kill switch converts single-provider risk into a procurement story, and Gemini Enterprise is compounding near 40 percent quarter on quarter on a multi-model Vertex marketplace.

Watch Amazon, medium — its CEO reportedly triggered the action against its largest investee weeks before Anthropic's targeted October listing. Q1 booked 16.8 billion pre-tax from that position, including a 12.3 billion non-cash revaluation. And watch Anthropic itself into that IPO window.

Highest conviction is the Alphabet long. The falsifier: Google Cloud growth decelerating below 30 percent year on year in the Q2 print, or no Gemini Enterprise acceleration cited. That settles by August 31.

We're also watching Alibaba at low conviction — over half of open-weight download share, but mindshare isn't revenue, and a US export mood makes Chinese weights a poor fit for the regulated workloads driving diversification.

The tape is the desk's scorecard, not advice.

Quick hits. A Trump order bars the Census Bureau and BEA from using statistical noise — experts say 2030 redistricting data plans must be completely redesigned.

The EFF warns H.R. 6028 would rush a disastrous overhaul of the US Copyright Office.

Apple migrated its TrueType hinting interpreter to Swift, and Renault detailed an electric motor that uses no rare earths.

A pancreatic tumour treatment may have revealed a cancer master switch, and areweguiyet surveys the state of building user interfaces in Rust.

And IEEE Spectrum argues the computer science degree isn't dead — file under skeptical, alongside one read on the shutdown titled simply, there is a shadow hanging over this Fable thing.

Our call: Fable 5, the safeguarded build, comes back online for customers within 60 days, while Mythos 5 stays dark under license — because Anthropic isn't fighting the order, it's building a record to fight the facts.

Proves us wrong if Fable 5 is still fully disabled for all customers on August 13, 2026 — that's when it settles.