Good morning. The Fable story has its villain — or at least, its named accuser. Reporting overnight from the WSJ, Verge and TechCrunch all converge on the same thread: Amazon CEO Andy Jassy personally raised the alarm with Treasury that triggered the export ban on Fable 5 and Mythos 5. Anthropic’s largest investor appears to have helped pull its largest models offline, and a Chinese lab released an open-weights frontier model the same afternoon. So that happened.
Jassy’s call to Treasury. The WSJ broke that Jassy flagged Amazon’s internal cyber research to Treasury Secretary Scott Bessent, after which David Sacks said Dario Amodei refused a government request to either patch the jailbreak or pull the model — and the export control followed. The Verge notes that security researcher Katie Moussouris reviewed Amazon’s paper and agrees with Anthropic that the vulnerabilities aren’t model-specific, with similar capabilities in GPT-5.5. Anthropic itself continues to push back publicly, telling The Verge the government provided only verbal evidence and no specifics.
The conflict-of-interest cloud. HN commenters spent the day picking at the obvious: Amazon owns more than 5% of Anthropic, runs Fable on Bedrock, and was reportedly furious about Anthropic’s new 30-day data retention rule that broke Bedrock’s compliance story. One read it as Jassy preferring nobody use Fable over customers using it through non-AWS channels; another compared it to SpaceX’s posture toward the current administration and suggested Anthropic simply hadn’t paid the right political dues. TechCrunch’s rundown is the cleanest summary if you’re catching up.
Zhipu ships GLM-5.2 into the gap. Zhipu released GLM-5.2 with a 1M context window, framed explicitly as a response to the Fable ban — the announcement went out at 5:21 PM Chinese time, the same hour Anthropic received the government letter, and cited “the sudden restriction of certain frontier models” as motivation. HN was appreciative but skeptical: benchmarks are missing, the release looks rushed, and several commenters asked whether “fully open” means open-weights or actually open-source. The broader sentiment, as one put it: “Open weight models are basically immune to that.”
Beijing forces Meta to unwind the Manus deal. In a mirror-image story from the other side, TechCrunch reports Meta is unwinding its $2B acquisition of Manus AI after a Chinese divestiture order on national security grounds, cutting Manus off from internal systems. The Manus founders are reportedly raising about $1B to take the company back, possibly via a Hong Kong listing. Beijing is now requiring approval for firms like Moonshot and ByteDance to accept US investment — the symmetry with this week’s US moves is hard to miss.
State AGs open an OpenAI investigation. A coalition of state attorneys general has subpoenaed OpenAI, led by New York, seeking documents on advertising, user data, sycophancy, and treatment of minors and seniors. It joins the Florida child-safety suit, copyright cases, and lawsuits tied to ChatGPT’s alleged role in user suicides — all while OpenAI’s confidential S-1 sits with the SEC. The NYT version drew HN comments from users who’d noticed ChatGPT inconsistently displaying crisis hotline numbers over the past year.
Munich court holds Google liable for AI Overviews. A Munich Regional Court ruled preliminarily that Google is directly liable for false statements generated by AI Overviews, after the feature falsely linked two publishers to scams by mixing their data with information about unrelated companies. The court found AI Overviews produces “independent, new, and substantial statements” rather than aggregating third-party content, so the usual search-engine safe harbors don’t apply. It’s a narrow German ruling for now, but the reasoning travels.
Why private inference isn’t enough for Siri. Matthew Green has a long post on Apple’s plan to power the next Siri with Gemini through Google’s Confidential Inference and Apple’s Private Cloud Compute. His argument: the cryptographic guarantees probably hold against Google and Apple seeing your data, but an agent reasoning over your emails, calendar and messages creates new failure modes that “private inference” doesn’t touch. Worth reading before WWDC details land.
That’s the week: a US lab loses its flagship to its own investor’s phone call, a Chinese lab loses its US buyer to Beijing, and a German court decides AI search isn’t search. Have a calm weekend.