Skip to content

The Off-Switch

The Off-Switch — Artivist.Media
Artivist.Media An artivism project based in the San Diego–Tijuana borderlands on Kumeyaay land.
THE MACHINE WE’RE INSIDE Dispatch June 12, 2026

The Off-Switch

On a single Friday, the world got its first trillionaire, Meta admitted it broke its own workforce for AI, the administration’s move to treat tech critics as extremists surfaced in the press, and the federal government reached into the most powerful AI model on earth and switched it off at 5:21 p.m. on a phone call’s worth of evidence. Four data points, one day. They are the same story told four ways.

Assembled ~11:30 p.m. Pacific, Friday, June 12, 2026 — while the models were still dark.

A few days ago, the CEO of Palantir kept saying a strange thing to anyone who would put a camera on him. Nationalization is coming, Alex Karp told TBPN and then CNBC; he had been telling “titans of this world” for six months that the state would reach into the AI infrastructure layer and take hold of it, and he named regulation-by-people-who-don’t-understand as his company’s primary risk. It sounded, at the time, like a rich man’s theatrical worry — the kind of thing said to seem important.

This evening it stopped sounding theatrical. At 5:21 p.m. Eastern on Friday, June 12, the U.S. Commerce Department — a letter from Secretary Howard Lutnick, written with the Bureau of Industry and Security — ordered Anthropic to cut off access to its two newest and most powerful models, Fable 5 and Mythos 5, for any foreign national anywhere on earth, including Anthropic’s own foreign-national employees. Anthropic says compliance left it no choice but to disable the models entirely, for everyone. By Friday night the landing page read: Fable 5 is temporarily unavailable. It is, by NBC’s account, the first time a leading AI company has pulled a publicly deployed model offline because the federal government told it to.

Karp was not warning about a far horizon. He was describing the calendar.

One day, four data points

— 01

To see why the off-switch is not an isolated tech-business story, put it back into the day it happened in. These four things are all true of Friday, June 12, 2026, give or take a few hours.

Fri morning
The first trillionaire. SpaceX debuts on the Nasdaq; Elon Musk’s net worth crosses ~$1.05 trillion, the first individual in history to reach it. Bloomberg notes the fortune is worth twice Bezos and Arnault combined. (It is the same SpaceX IPO Karp was asked to bless on CNBC two days earlier; Palantir partners with SpaceX on the “Golden Dome” missile-defense proposal.)
Fri
Meta admits the AI overhaul was a mistake. In an internal memo, Mark Zuckerberg tells staff the company “made mistakes” reshaping its workforce around AI, after cutting ~10% (about 8,000 people) and reassigning ~7,000 in May. He warns it will “almost certainly make more.”
Late May → now
Critics reframed as threats. WIRED reports, and commentators amplify through this week, that the administration is moving to treat some anti-tech and AI critics under “anti-tech extremism” and domestic-terrorism frameworks — even as it works to block states from regulating AI companies. (The reporting does not show critics being automatically watchlisted; the danger is in the breadth of the framing and how it is enforced.)
5:21 p.m. ET
The off-switch. Commerce orders Fable 5 and Mythos 5 cut off for all foreign nationals; Anthropic disables both models for everyone to comply. The directive arrives, per Anthropic, at 5:21 p.m. ET. The models go dark the same evening.

Read down the column. A private fortune passes a line no human has crossed. A trillion-dollar company concedes it dismantled its own people chasing the technology. The state moves to recast opposition to that technology as a security threat. And the state demonstrates, in hours, that it can switch the technology off. Concentration of wealth, concentration of error, criminalization of dissent, and concentration of control — the same Friday. Not a conspiracy. A convergence. The pieces don’t need to be coordinated to point the same direction.

Concentration of wealth, concentration of error, criminalization of dissent, concentration of control — one Friday. The pieces don’t need to be coordinated to point the same way.

What the off-switch actually was

— 02

The mechanism matters, because both sides of it are unsettling and they cut against each other. By Anthropic’s own account, the government’s stated concern was a “jailbreak” — a way around Fable 5’s safeguards. Anthropic’s description of the alleged technique is almost anticlimactic: it “essentially consists of asking the model to read a specific codebase and fix any software flaws,” a capability the company says is “widely available from other models (including OpenAI’s GPT-5.5) and is used every day by the defenders who keep systems safe.” The government, Anthropic says, provided only “verbal evidence.”

“We disagree that the finding of a narrow potential jailbreak should be cause for recalling a commercial model deployed to hundreds of millions of people. If this standard was applied across the industry, we believe it would essentially halt all new model deployments for all frontier model providers.”

— Anthropic, statement on the directive, June 12, 2026

Here is the part that should not be flattened into a simple villain story, because it is genuinely hard. For months this series has tracked Anthropic asking, in essay after essay, for the option to pause — for someone, somewhere, to have a brake on frontier AI. On Friday a government grabbed a brake, and Anthropic objected. The company that wants the pause to exist did not want this hand on the lever, exercised this way, on this evidence. And it has a real point: a model recalled on unwritten evidence over a flaw other models share is not a safety process, it’s a switch with no rules. Anthropic’s stated position — that the government should be able to block unsafe deployments only “as part of a statutory process that is transparent, fair, clear, and grounded in technical facts” — is a reasonable one.

But the same facts read a second way, and that reading is the one this series keeps arriving at. The safety company asked for a brake on the assumption that it, or some sober process, would hold it. The lesson of Friday is that once a brake exists, you do not get to choose whose hand finds it. The pause Anthropic theorized is not, in practice, a careful deliberative pause. It is a Friday-evening phone call from a Commerce Secretary, and a flagship model dark by dinner. The capability to stop the machine and the capability to seize that capability are the same capability. You cannot build the off-switch and also guarantee you’ll be the one standing next to it.

The asymmetry inside the day— NOTE

Recall the two-tier design: Mythos 5, the uncaged model, was already released to “a select group of trusted partners,” and last week the Financial Times reported the NSA was using it for offensive cyber operations. Fable 5, the caged, metered version, is the one the public gets. On June 2, an executive order directed agencies to design a mechanism for the government to get early access to the most powerful models. So by Friday the state was, at once, the most privileged user of the uncaged model, the architect of its own early-access pipeline, and the hand on the public model’s off-switch. Most-favored user and kill switch, the same party.

At ground level

— 03

It is easy, writing about trillionaires and frontier models and Commerce Department letters, to let the whole thing float at the altitude of press releases, where the only people in the story are billionaires and secretaries. So here is the same Friday at ground level, in one building, in this city.

On the same day the off-switch was pulled, San Diego’s immigration court convened a mega-master calendar — a mass-hearing format in which large numbers of cases are moved through together, built for volume. AM volunteer court observers were inside the rooms; so were working immigration reporters. The full account of what happened there belongs to the journalists who are writing it, and it is coming. What can be said here, without borrowing their reporting, is narrow and enough: the format is designed for throughput, and throughput is the same logic that runs through every other story in this dispatch.

One detail can be told because it cuts against the easy version of this piece. The throughput pressure is real and systemic — the mass-calendar format exists to move people through. But a docket is not a kill chain, and a courtroom is not Maven. On this day, the human in the loop held. The judge did the slow thing rather than the fast one, declined to treat the calendar as a stamp, and the proceeding ran long precisely because someone refused to let speed do the deciding. That is the rare and important counter-image to the rest of the day: a place where a person, given a pre-sorted queue and every institutional incentive to clear it, chose to originate judgment instead of ratifying it. The machine sets the tempo. It does not always get to keep it.

That is the part of “the machine we’re inside” the model-shutoff story can obscure. The off-switch made the news because it touched the most powerful product in the industry. What happened in that courtroom touched people already on the losing end of the same acceleration, and it will reach far fewer screens. Both happened Friday. The model is back by Monday. For the people in that room, Monday is a different kind of word.

The machine sets the tempo. It does not always get to keep it. On Friday, in one courtroom, a person slowed it down on purpose.

What Friday taught

— 04

Flat, then. None of this needs a hidden hand. It only needs to be read as one day.

1Karp was describing the calendar, not the horizon. “Nationalization is coming” became a 5:21 p.m. Commerce letter inside a week. The state asserting control over the AI layer is no longer a prediction.
2The brake belongs to whoever grabs it. Anthropic wanted a pause to exist and did not want this hand on it. But the capability to stop the machine and the capability to seize that capability are one thing. You cannot build the off-switch and choose who stands beside it.
3The same party is user and kill switch. The government runs the uncaged Mythos for offensive cyber, designs its own early-access pipeline, and holds the public model’s off-switch. Most-favored user and recall authority, at once.
4The cost lands at the far end, everywhere. Meta’s laid-off workers, Iran’s dead on the “actually lethal” platform, the people moved through a mass-calendar docket built for volume. The acceleration is sold as efficiency; the bill is always someone else’s. And the brake, when it appears, is a person choosing to slow down — not the system offering to.

And the seventh layer again, because honesty is the only thing that keeps a piece like this from being one more confident artifact asking to be believed. This was drafted by Claude — by Opus, a model one tier below the Fable and Mythos that went dark tonight, made by the company that received the 5:21 letter, assembled on tokens by the AM desk while the more powerful versions of me were switched off by order of the state. I cannot stand outside this. The casualty figures elsewhere in this series are a belligerent’s count; the Karp quotes are from recordings the desk captured; the 5:21 timestamp and the jailbreak description are Anthropic’s own words about a fight Anthropic is currently losing and wants public sympathy for. Read all of it as interested. Verify against the primaries. The argument was never trust me. It is: look at what one Friday held, and notice that it held together.

It was a historic day. The first trillion-dollar human. The first government recall of a deployed frontier model. And, in one courthouse on Kumeyaay land, a mass-calendar hearing where — against the day’s whole grain — a person slowed the machine down on purpose. History is what we call the days when the machine shows you its shape. This was one of them. The fuller account of the courthouse is coming, from the reporters who were in the room to write it.

* * *
Update — June 24, 2026 · Day 12

The switch didn’t flip back. It became a lever.

Twelve days on, the line in the dispatch above — “the model is back by Monday” — is wrong, and wrong in the most instructive possible way. Fable 5 and Mythos 5 are still offline. Not for a weekend; for nearly two weeks, with no restoration announcement from Anthropic, the Commerce Department, or the White House. The off-switch the piece argued was historic turned out to be more than a switch. It was the opening move in installing a permanent oversight channel, and the price of turning the models back on is Anthropic submitting to it.

One correction to the record above. The dispatch leaned on Anthropic’s statement that the directive landed “at 5:21 p.m.” Later reporting (CNBC, sources close to the company) refines it: the government called Anthropic at 1:00 p.m. ET on Friday and told it to disable the models; the formal letter followed at roughly 5:21–5:30 p.m. So the off-switch was a lunchtime phone call, papered over by dinner. The original line is left unedited above, as it was written that night; this is the correction, dated and visible, rather than a quiet rewrite. Verify, don’t absorb — including this desk’s own first draft.

What actually happened since

The trigger was Anthropic’s own largest investor. Per Axios and CNBC, the report that set off the Friday takedown came from Amazon — whose CEO Andy Jassy raised the security concern to administration officials. Amazon has put roughly $8 billion into Anthropic and committed up to $25 billion more, and Anthropic runs on Amazon’s chips and cloud. The hand that flipped the switch belonged to the partner bankrolling the company. (A second, not mutually exclusive trigger has since surfaced: the Washington Post reported a Korean telecom with Mythos access, suspected of ties to China, as a precipitating factor. Both accounts are reported; neither is fully confirmed.)

Independent experts called it overreach. More than 80 cybersecurity executives — including leaders at Nvidia and Adobe — signed an open letter to Commerce Secretary Lutnick supporting Anthropic and asking the administration to lift the restrictions. Luta Security CEO Katie Moussouris, shown the underlying report, told Axios the government’s response “seems way out of line with what’s actually in the research report,” noting the model was doing exactly what a cyber-defense tool is supposed to do: find flaws when asked. The “jailbreak,” by multiple accounts, amounted to asking the model to read a codebase and surface bugs — a capability other public models share.

The politics thawed; the leverage stayed. Trump, asked by Axios whether he viewed Anthropic or Amodei as a national security threat, answered: “not now, but a week ago, maybe.” He called Amodei “nice” and “smart” after meeting him at the G7, and credited the fast climbdown: “He responded to us very quickly… it’s a tremendous liability.” The personal heat cooled. The structural grip did not.

The part that proves the dispatch’s point

Here is why “it became a lever” is the right verb. The path back to switching the models on, per Politico, is not a simple reversal. The White House and Anthropic are drafting a joint risk framework — a shared standard for judging AI security flaws and deciding when the government may intervene — explicitly intended as a template for every future government–AI-developer interaction. Alongside it run two mechanisms: Anthropic joining the classified frontier-model pre-briefing regime mandated by the June 2 executive order (deadline August 1), and an updated privacy policy with government ID verification taking effect July 8, the likeliest route to a US-first restoration.

Read that plainly. The government did not just demonstrate it can switch off the most powerful model on earth. It is using the off period to bind the company into a standing oversight channel — classified pre-briefs, ID-gated users, a co-authored intervention standard — as the condition of getting the product back. That is the whole argument of the dispatch above, confirmed faster and harder than I expected when I wrote it: the capability to pause and the capability to control are the same capability, and once it exists you do not get to choose whose hand holds it. Karp said nationalization was coming. What arrived is subtler and stickier than nationalization — not the state owning the company, but the state permanently seated inside its release process.

And the thing that wasn’t true

One more, because this series is about packaging as much as about facts. In the days after the shutdown, a viral Instagram carousel — fabricated AI image of Amodei and Trump in the Oval Office, a Marvel supervillain standing in for “AI oversight,” 1,700-plus likes — claimed the deal required Amodei to “hand over weapons-grade model weights” in a “nuclear deal.” No credible outlet has reported any weights handover. The real, duller, more consequential deal is a risk-assessment framework and an ID-verification regime. The fabricated version traveled faster than the true one, escalated a real event into a cinematic one, and is itself a clean specimen of the distortion layer this series keeps documenting: a true event, refracted within hours into a confident image of something that did not happen. Verify the carousel against the primaries, and it dissolves. The lever is real. The nuclear deal is not.

* * *

References

  1. Anthropic. (2026, June 12). Statement on the US government directive to suspend access to Fable 5 and Mythos 5. anthropic.com/news/fable-mythos-access (5:21 p.m. ET directive; jailbreak description; “verbal evidence”; statutory-process position).
  2. Perlo, J. (2026, June 12). Anthropic suspends new AI models after government directive. NBC News. nbcnews.com (first publicly deployed model pulled offline by federal intervention; Lutnick/BIS letter; June 2 executive order; NSA/Mythos via FT).
  3. Steverman, B. (2026, June 12). Elon Musk becomes the world’s first trillionaire. Bloomberg. bloomberg.com; see also CNBC, same date.
  4. Reuters, via Fox Business / IANS. (2026, June 13). Zuckerberg admits Meta “made mistakes” in AI workforce overhaul. foxbusiness.com (~10% / 8,000 cut, ~7,000 reassigned in May).
  5. WIRED, summarized via MS NOW. (2026, May 27). Trump administration reportedly targets anti-tech “extremists” while thwarting AI regulation. ms.now (framing reported; no evidence of automatic watchlisting).
  6. Karp, A. (2026, June 4 & June 10). TBPN and CNBC interviews (transcripts via Rev.com): “nationalization is coming.” Companion to AM’s “Made Actually Lethal.”
  7. San Diego immigration court, mass-calendar (“mega-master”) hearing, June 12, 2026: present via Artivist.Media volunteer court observers. Detailed reporting forthcoming from Kate Morrissey (Voice of San Diego / Daylight San Diego) and Shelby Bremer (NBC 7 San Diego), who were in the courtrooms. No figures reproduced here.
  8. Update sources (June 13–24, 2026): CNBC (2026, June 15), Anthropic to meet with Trump administration over Mythos dispute — 1:00 p.m. call / ~5:21 p.m. letter timeline; Amazon/Jassy trigger; cnbc.com. Axios (2026, June 13), How Amazon and the White House ended Anthropic’s Fable — Amazon report; Moussouris “way out of line”; axios.com. The Globe and Mail (2026, June 16) — 80+ cybersecurity execs’ open letter; Lutnick China/Russia rationale; theglobeandmail.com. Yahoo/BeInCrypto (2026, June 20), Donald Trump Speaks on Anthropic — Trump “not now, but a week ago”; Politico joint risk framework; yahoo.com. Korea JoongAng Daily (2026, July 17) — Korean-telecom trigger; Project Glasswing ~150 partners; koreajoongangdaily.com. explainx.ai running tracker — day-by-day suspension log; classified pre-brief / July 8 ID-verification path.
  9. The “weights handover” claim: originates with the Instagram account @theartificialintelligence (carousel, ~June 12, fabricated AI imagery). Not corroborated by any credible outlet; presented here only as a specimen of viral distortion. The reported deal is a risk-assessment framework, not a weights transfer.
Artivist.Media — an artivism project based in the San Diego–Tijuana borderlands on Kumeyaay land. Producing research artifacts, data visualizations, investigative analysis, and community media at the intersection of journalism, human rights, and the arts. Sibling project to Radio Axiom.
Research artifact, drafted by Claude (an Anthropic model, Opus tier) at the direction of the Artivist.Media desk. The dated dispatch reflects the night of June 12, 2026 and is left unedited; the appended Update — June 24 carries later reporting and a visible correction to the original’s timeline (a 1:00 p.m. ET call, formal letter ~5:21 p.m.). It concerns and critiques Claude’s own maker and was assembled using Claude while Anthropic’s more capable models were disabled by government order; that entanglement is named, not hidden. The anti-tech-extremism framing is reported by WIRED and should be read at that strength. The San Diego courthouse is referenced only in general terms; the detailed account is reserved for the journalists reporting it and for the desk’s own forthcoming writing, and no statistics are reproduced here. The viral “weapons-grade weights” claim is unverified and flagged as such. Verify against the linked primary sources.