In partnership with
Sponsor:
HubSpot AEO
Picture this. A buyer opens ChatGPT and asks for a recommendation in your category. Your competitor's name comes up. Yours doesn't. And that buyer never makes it to your website.
That's happening right now in markets everywhere. And most teams don't know it's happening because it never shows up in their analytics.
HubSpot AEO shows you exactly where your brand stands in AI search, where competitors are getting recommended instead of you, and tells you specifically what to fix. No expertise needed.
Try it free for 28 days. Just $50 a month after.
iPrompt
THE AI NEWSLETTER THAT TURNS NEWS INTO ACTION
ISSUE #140 WEDNESDAY · 17 JUNE 2026
THE HOOK
Last Wednesday we told you to get on the right side of the gate before it mattered. Six days later it mattered — from a direction we didn't price. On 12 June, by its own statement, Anthropic switched off Fable 5 and Mythos 5 — the most capable models it had ever shipped — to comply with a US government export-control order. Not throttled. Not metered. Switched off, worldwide, for every non-US citizen. The disqualifier wasn't your budget. It was your nationality, and it changed overnight.
AI NEWS ROUNDUP
This week in AI
1 The government took Fable 5 off the board — and Anthropic confirms it. Anthropic's own statement says it received an export-control directive on 12 June at 5:21pm ET suspending Fable 5 and Mythos 5 for any foreign national, in or outside the US. It can't verify nationality in real time, so it disabled both for everyone. The trigger, per an administration official cited by Axios, was another firm's jailbreak claim — which Anthropic calls narrow and already replicable on other public models. Reporting frames it as the first export control aimed at a model, not a chip. Anthropic →
2 China answered in 24 hours. One day after the ban, Z.ai shipped GLM-5.2 — open weights under MIT, a million-token context, running in Claude Code with a config change — and explicitly cited the shutdown as proof US models can't be relied on abroad. Zhipu's stock jumped 33%. The "risky" open model is now the one you can actually run. Gigazine →
3 Defenders revolted — and the "jailbreak" may have been three words. Nearly 150 security leaders, organised by ex-Facebook CSO Alex Stamos, signed an open letter to Commerce urging reversal: pulling the best tools from defenders while adversaries keep building is, they argue, sabotage, not safety. Researcher Katie Moussouris, the one outside expert to read the underlying paper, reports the trigger was essentially asking the model to "fix this code." Axios →
4 Five days on, still dark — with no committed date. Both models remain suspended; Anthropic says only that it's working to restore access. Per the company, the claude-fable-5 API string now errors out and the recommended fallback is claude-opus-4-8. A third-party "48-hour" rumour spread Tuesday — Anthropic has confirmed nothing, and prediction markets are split on a near-term return. Polymarket →
5 The contrast of the week: capital scaled up as control clamped down. The same Friday Commerce switched off a frontier model, SpaceX — betting on AI data centres in orbit — closed up ~19% in the largest IPO on record. Build as fast as money allows; lose your access overnight to a decision that isn't your vendor's. Both true at once. CNBC →
OUR ANGLE
🔭 We had the gate right and the gatekeeper wrong. The call: last week we argued capability had stopped being the product and access was the new scarce good — verify, then permit. That was right. The miss: we assumed the lab held the key and you bought your way in with a trust file. The state held a bigger one. It didn't raise the bar for the vetted — it removed the model from under everyone who isn't a US citizen, in an afternoon. Vetting didn't buy you in. Nationality locked you out. What I infer (my read, not reported fact): access risk now has two independent sources. One is the lab's classifier — a price-and-permission limit you can plan around. The other is the state's export switch — a binary, retroactive kill that no enterprise contract survived. GLM-5.2 is the tell: the hour the closed American model went dark, an open-weights one ran straight through the wall beside it. The control only bites where the model is closed. My forecast — against the herd: most are reading this as a one-off diplomatic snafu fixed in a fortnight. I think it's the founding precedent. By the end of 2026, at least one more publicly deployed frontier model is restricted or geo-fenced by a government on national-security grounds — and "AI-provider continuity" becomes a named board-level risk line, the way cloud-region failover already is. The question stopped being which model is best. It's now: what do you run when your best one is switched off by someone who isn't your vendor? |
THE THREE SPECIALS
Do · Use · Understand
🎯 PROMPT OF THE WEEK
The Continuity Stress Test
Last week's prompt asked whether the frontier tier was worth paying for. This one asks the question 12 June made urgent: what actually breaks if your primary model vanishes at 5pm with no warning? Paste this into the model you depend on most. It maps your single points of failure and writes the fallback runbook before you need it.
You are my AI continuity auditor. Assume my primary model is switched off in the next hour — vendor outage, a pricing change, or a government order. No warning. MY SETUP: - Primary model: [name it] - What I use it for: [list your 3 most important workflows and what 'broken' looks like for each] - Where it's wired in: [apps, APIs, automations] Produce: 1. A single-point-of-failure map — which workflows die instantly, which degrade, which are fine. 2. The fallback for each: a specific named model I can switch to today, and what I lose by doing so. 3. The 15-minute switch plan — exact steps, in order, to get my top workflow running on the fallback. 4. One thing I should set up NOW so the next outage costs me minutes, not days. Flag every assumption you make in [brackets]. |
Why it works: an outage you've rehearsed is an inconvenience; one you haven't is a crisis. Logicalis's 2026 Global CIO Report — over 1,000 CIOs surveyed — found 16% have no continuity plan at all for a key AI provider going dark. That was a footnote in March. The 12 June shutdown turned it into a live fire drill. This prompt makes you the five-in-six that's ready. Where to be careful: a fallback you've never run isn't a fallback — it's a hope. Once the model names your switch target, actually run your top workflow on it once this week. Paper plans fail at 5pm. Tested on: Opus 4.8 and GPT-5.5, on two of our own production pipelines. Both produced usable switch plans; Opus was sharper on naming the exact degradation per workflow. |
🛠️ TOOL OF THE WEEK OpenRouter One API key, one bill, ~hundreds of models behind it — flip your fallback with a string, not a rebuild. ★★★★½ / 5 Use if: you build on AI and 12 June just taught you that hard-coding one vendor's model string is a liability. OpenRouter puts a routing layer between you and the labs, so swapping Fable for Opus, GPT-5.5 or an open model like GLM-5.2 is a one-line change, not a sprint. Skip if: you only ever use one chat app on one plan and never touch an API — the routing layer buys you nothing. Describe it to a colleague: "It's the universal adapter for AI models — one socket, swap the brain behind it whenever you need to." Best use case: wire your most critical workflow through it now, with a primary and a named fallback already configured. Then the next government letter or outage costs you a config edit, not a weekend. |
💡 TIP OF THE WEEK The real risk isn't downtime. It's correlation. Where it doesn't apply: if AI is a convenience you could drop for a week with no real cost, skip this — over-engineering a hobby is its own waste. This is for the workflow that genuinely matters. The principle: the reason 12 June hurt thousands of teams at once is that they shared a single dependency that failed at the same instant, for the same external reason none of them controlled. That's correlated failure — and it's the risk that doesn't show up in any vendor's uptime number, because the vendor was fine. The model was switched off above its head. A second model from a different provider only helps if its failure isn't correlated with your first: different company, ideally different jurisdiction. Two US closed models share the same kill-switch; a US model plus a self-hostable open-weights one do not. How to apply it: for any workflow you'd panic to lose, ask not "is my provider reliable?" but "what single event takes my primary AND my backup down together?" If you can name one — same vendor, same country, same regulator — you don't have a backup, you have a duplicate. Pro move: once a quarter, run a "chaos hour" — pretend your primary is gone and do real work on the uncorrelated fallback for sixty minutes. You'll find the broken prompt and the missing integration while it's cheap to fix, not at 5pm on a Friday. |
YOUR MOVE
One action. Reply by Friday.
This week in three lines:
Anthropic confirms it switched off its two most capable models on 12 June to comply with a US export order — reported as the first aimed at a model, not a chip.
We had the gate right and the gatekeeper wrong: what locked you out wasn't price or vetting but nationality, decided in Washington and applied in an afternoon.
An open-weights model ran through the wall a day later — so your real protection isn't a bigger budget, it's a fallback whose failure isn't tied to your first one's.
Your one action: run the Continuity Stress Test on your single most important AI workflow today, while the lesson is fresh. Then reply with one line: the workflow that would hurt most to lose. I read every response — and the answers tell me what next week's issue owes you.
Optional, after you've replied: the deep dive on the kill-switch era and what AI-provider continuity now means for your board. The reply comes first.
R. Lauritsen
P.S. Six days ago I told you to get vetted before the gate mattered. It mattered faster than I thought, and from a direction I didn't fully price — being known wasn't the protection; an uncorrelated fallback is. I'd rather be useful than right, so hit reply and tell me where your stack is exposed. Then, if you want the longer argument, the deep dive makes the case that continuity is the new uptime.
Forward iPrompt For the colleague who still thinks the only AI risk is the bill. |
Fix that. Live. With Clay + HubSpot.
Defining your ICP on vibes is a pipeline killer. In Build Your GTM Alpha, Clay + HubSpot for Startups walk you through a live build. Real prospect list. Real enrichment. Real outreach sequence. You don't leave with a plan. You leave with outbound running. June 18. 11am ET / 4pm GMT.