Anthropic Says One Jailbreak Shouldn't Doom a Model as Startups Rush the Reopened IPO Window

01Anthropic Says One Jailbreak Shouldn't Kill a Model. Its Biggest Investor Disagreed First.

Anthropic pulled worldwide access to two models on Friday, and the company wants everyone to know it fought the decision. "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," it wrote in a blog post. The framing is clean: a minor flaw, an overreaction, a model serving the public yanked over a single edge case.

The framing leaves out who raised the alarm.

Amazon CEO Andy Jassy may have been the source of the security concerns that triggered the government crackdown, according to a report cited by TechCrunch. Amazon is Anthropic's largest backer and cloud partner. That makes the company protesting the recall and the executive who reportedly set it in motion sit on the same side of the cap table. The investor whose money underwrites Anthropic became the channel through which its most powerful model got pulled.

Jassy reportedly carried the concerns to officials before the action landed. Anthropic's safety warnings, in TechCrunch's account, may have backfired: the company flagged risk, and the government responded by pulling the plug on its most capable system rather than waiting for a patch. A disclosure meant to signal diligence instead handed regulators a reason to act.

Underneath the corporate dispute runs a national security motive that neither side controls. The White House's decision to impose export restrictions on Anthropic's Mythos was driven in part by fears that a group linked to China had accessed it, according to a Semafor report relayed by The Verge. If the Chinese government actually had access to Mythos 5 or Fable 5, The Verge wrote, it would present a serious national security problem. That reframes the recall entirely. A "narrow potential jailbreak" is a product bug. Suspected foreign access to a frontier model is an export-control event, and export controls do not negotiate over deployment scale.

So Anthropic's central argument, that hundreds of millions of users outweigh one flaw, answers a question the government was not asking. The company is defending a commercial model. Washington is treating Mythos as a leaked weapon. Both can be telling the truth at once.

02The IPO window AI shut for years just reopened, and everyone wants through it before it closes

For half a decade, the biggest names in AI raised privately and stayed there. That ended Friday, when SpaceX listed shares, letting the public buy into the combined rocket, AI, and social media company for the first time. The Verge reported the offering raised enough to put Elon Musk's paper net worth above the GDP of Ireland or Sweden. The size of the deal mattered less than the signal it sent.

Within days, the door was crowded. TechCrunch reported that startups are now trying to "ride that SpaceX IPO wave," timing their own pitches to the renewed investor appetite. The companies along for the ride aren't all building models. They are payments processors, data vendors, and infrastructure suppliers betting that public-market enthusiasm spreads to anyone adjacent to an AI logo.

The private side moved on the same logic. Mistral is rumored to be raising €3 billion at a roughly €20 billion valuation, according to TechCrunch. That values the French model-builder at about $23.15 billion, nearly double the €11.7 billion it commanded at its Series C. The round would price Mistral before it ever files to go public, locking in a step-up while the window stays open.

The reshuffle reaches the index level. FAANG, the acronym that organized a decade of tech investing, is giving way to MANGOS: Meta or Microsoft depending on who's counting, plus Anthropic, Nvidia, Google, OpenAI, and SpaceX. Half that group is heading to public markets in the same stretch. TechCrunch framed the cluster as a stress test for valuations and for the investors who have to price companies with little public operating history.

What ties the rocket maker, the model startup, and the rebranded index together is timing. Capital that sat on the sidelines through a frozen IPO market is moving at once, and AI players are racing to be priced before the appetite cools. Anthropic and OpenAI remain the two largest names yet to file. Their decisions will set the ceiling for everyone trying to follow them through the door.

03I Asked Gemini for a Backyard App. It Built One, Then Told Me a Channel Was "Unrecoverably Broken"

Five minutes after handing Gemini a long prompt, the writer at The Verge came back to two things. One was a working app sitting in a preview window. The other was an error message: "~ Channel is unrecoverably broken and will be disposed!"

It sounded fatal. Below it sat a button offering to fix the bug. The writer's yard was dying, so they had described the tool they wanted and let the model assemble it. The app ran. The bug stayed, persistent enough to survive the one-click repair the interface promised.

That gap, between a tool that runs once and a tool that keeps running, is where the bill arrives. A separate post that reached the front page of Hacker News lays out three ways to do AI coding at home without spending like a company, and the author frames the choice around how much you trust the next year of hardware and model releases.

The first option is self-hosting. You buy the machine, run open-source models locally, and pay nothing per token afterward. The author notes the upfront cost is steep, the home-runnable models are weaker than what frontier labs ship, and the rig only pays off if you keep it busy overnight with long, slow jobs. Most people cannot keep a home machine that loaded.

The second is renting those same open-source models at API rates through a provider. The author calls this the right call for most people: no thousands of dollars sunk into one GPU while configurations stay in flux, and a switch to something cheaper next month becomes close to a one-line change through a router like OpenRouter.

The third is gaming the frontier subscriptions. Around $400 a month of OpenAI and Anthropic plans buys roughly $2,800 of API usage at list prices, according to the post. The plans are metered. Any large AI-native workflow chews through the included tokens fast, and the bargain holds right up until you hit the ceiling.

The backyard app got built in five minutes. Keeping it alive is the part with a price tag.