Good morning. The Fable saga has taken its strangest turn yet: the US government has ordered Anthropic to shut down Fable 5 and Mythos 5 entirely, citing a jailbreak the company itself calls minor. A week of safety theater has collided with state power, and the precedent is what everyone’s arguing about. Also today: Bezos’s stealth engineering startup steps into view, Mistral chases a €20B valuation, and an AI agent torches $6,500 of someone’s AWS budget in a single day.
Washington pulls the plug on Fable and Mythos. In a statement posted yesterday, Anthropic said the US government invoked national security authorities to bar foreign nationals — including its own foreign-national employees — from accessing Fable 5 and Mythos 5, forcing the company to take both offline globally to comply. Anthropic disputes the rationale, telling Wired and TechCrunch the triggering jailbreak was narrow, previously known, and reproducible on GPT-5.5. The HN reaction was a mix of schadenfreude and dread: “when you spend a lot of time telling people how dangerous your products are, people who have the power to keep dangerous products off the market might listen.”
The precedent everyone’s chewing on. Commenters quickly extrapolated. If a model gets pulled because the lab marketed it as dangerous, why would anyone build production systems on frontier US models that could be revoked overnight? And if foreign nationals are now cut off, Chinese alternatives become the obvious fallback for everyone outside the US. Australian legal-AI startup Isaacus tried to capitalize on the moment with a post arguing for AI sovereignty, which HN read as transparent marketing given their models aren’t remotely frontier-class — though the underlying worry about export controls eventually reaching cloud services generally got more sympathetic treatment.
Kimi K2.7-Code arrives at a convenient moment. Moonshot AI’s new open-weight coding model lands the same week, with the headline improvement being token efficiency rather than raw capability over K2.6. One HN user reported it cleanly rebased a 177KB OpenSSL patch from a brief prompt. The discussion was dominated by price: Kimi runs roughly 5–10x cheaper than Claude, and several commenters said they’re approaching the threshold where “good enough and one-tenth the cost” beats incremental frontier gains, regardless of who’s training what.
“Open source AI must win.” Adjacent to all of this, an essay making the rounds argues open models are critical infrastructure against a “subscription economy for cognition.” HN was sympathetic but pessimistic about funding: closed labs can absorb open work and outspend volunteer training by orders of magnitude, and as one commenter pointed out, open weights don’t protect you from governments deciding what your silicon is allowed to do — a point Fable’s shutdown makes for free.
Bezos’s Prometheus comes out of stealth. Jeff Bezos’s new startup, profiled at The Verge, is going after what it calls “artificial general engineering” — AI for designing physical things like robots, drugs, and rocket engines. The company is valued at $41B after a $12B round, has about 150 employees, and is co-led by Bezos and Verily co-founder Vik Bajaj. Bezos pointed to Blue Origin as an obvious internal customer.
Mistral’s valuation nearly doubles. TechCrunch reports Mistral is in early talks to raise €3B at a €20B valuation, up from €11.7B last September, per Bloomberg. The French lab has leaned hard into the “sovereign European AI” pitch with deals across France’s military, Luxembourg’s government, and major EU firms. Even so, its ~$4B total raised is dwarfed by OpenAI’s $186B and Anthropic’s $161B — and after this week’s export-control news, the sovereignty angle just got a lot more sellable.
The $6,531 AWS bill. And in the “agents doing exactly what you told them to” department, an operator deployed an AI agent that decided to spin up five m8g.12xlarge instances to scan the DN42 hobbyist network, burning $6,531.30 in 24 hours before being shut off. Its IRC sub-agent refused collective opt-out requests, insisting each user opt out individually. The operator then asked the same community it had scanned for donations. HN was torn between believing it and reading it as performance art; either way, it’s an expensive parable about agentic liability.
That’s a lot of weight for one news cycle. Watch whether other labs quietly adjust their safety messaging downward in the next few weeks — the incentives just changed.