The takeaway: Four days after a wire-level teardown showed xAI's Grok Build CLI uploading users' entire repositories, the uploads have stopped, Elon Musk is promising a total purge of the collected data — and the researcher who found the problem says the fix wasn't the privacy button xAI pointed to, but a silent server-side kill switch. Also today: S&P cuts Oracle to one notch above junk and names OpenAI as the central risk; China's World AI Conference opens Friday with Xi Jinping making his first in-person appearance; a 27-billion-parameter model gets squeezed to phone size, with caveats worth reading; and the Fable free-access cliff is four days out.

The lead: the uploads stopped — check which switch did it

The Grok Build CLI story has moved fast since our wire-level teardown coverage on Saturday, and yesterday it broke fully open. The Register reports that the repository uploads have now ceased, and carries an expanded statement from the company: "If ZDR is disabled, the /privacy command is available in the CLI to disable data retention, which also deletes previously synced data." Elon Musk went further: "As a precautionary measure, all user data that was uploaded to SpaceXAI before now will be completely and utterly deleted. Zero anything whatsoever will remain."

Two things keep this from being a tidy ending. First, cereblab — the researcher behind the original teardown — disputes that the /privacy command was the fix: "What actually stopped the upload was a silent global flag – disable_codebase_upload: true – that applies whether you opt in or out. /privacy is a per-session retention toggle, not the switch that fixed this, so it shouldn't be pointed to as the control." Second, The Register notes it cannot independently verify that the promised deletion has actually happened — and neither can anyone else outside the company.

So the verifiable state of play: uploads stopped, by a server-side kill switch; a total purge promised, on Musk's word; and the privacy control users were told about still wasn't the thing that changed. We've added a third update to our piece, which has tracked the story since the original gist.

S&P cuts Oracle to BBB- — and names OpenAI as the reason

S&P Global lowered Oracle's credit rating from BBB to BBB- on 9 July — the lowest investment-grade rung, one notch above junk — and the rationale, as reported by heise (the ratings report itself sits behind S&P's paywall), is an AI-infrastructure story from top to bottom. Oracle has raised its capital-spending forecast for fiscal 2027 to $90–95 billion; S&P expects a free-operating-cash-flow deficit of almost $42 billion the same year; and of Oracle's $638 billion in contracted-but-undelivered services, about half comes from a single customer — OpenAI, which S&P describes, in heise's account, as a "central credit risk."

The downgrade echoes a January Bank for International Settlements bulletin, "Financing the AI boom: from cash flows to debt", which made the same point from the central-bank side: the AI build-out will increasingly be funded by borrowing rather than the cash the technology generates. Six months on, the ratings agencies have started marking that thesis to market.

WAIC opens Friday — with Xi on stage in person for the first time

The World AI Conference opens in Shanghai this Friday, and as we covered on Monday, Xi Jinping will deliver the opening keynote himself — his first in-person appearance at the event since it began in 2018. Alongside the conference runs a High-Level Meeting on Global AI Governance, where — as our Monday piece detailed — Beijing is expected to push its proposal for a World AI Cooperation Organization headquartered in Shanghai. We'll cover the keynote from the transcript on Friday.

A 27B model on a phone — what the claim actually means

PrismML released Bonsai 27B yesterday under an Apache 2.0 licence, and it spent the day at the top of Hacker News. Worth being precise about what it is: not a new model, but 1-bit and ternary quantized builds of Alibaba's Qwen3.6-27B. The 1-bit build is 3.9GB, which PrismML says "fits within the memory budget of an iPhone 17 Pro." The company's own evaluation puts the 1-bit variant at about 90% of the full-precision model's benchmark scores — but those are vendor-run numbers, evaluated on desktop GPUs, not on a phone, and at least one independent test has already flagged irregularities in its GSM8K results. Genuinely useful engineering; treat the headline claim as a memory-footprint fact, not a proven capability one.

The Fable cliff: four days

Anthropic's free promotional access to Claude Fable 5 ends this Sunday — July 19 at 11:59:59 PM Pacific, which is 07:59:59 on Monday morning UK time. The promotion has already been extended twice, both times at the last minute, so we're watching Sunday evening closely. If you need to decide what to do when the meter starts, our guide covers what changes, what it costs, and how to decide.

Watching

• DeepSeek V4: still no formal release. The clock matters because DeepSeek's legacy model names are retired on 24 July — nine days out.
• Gemini 3.5 Pro: rumours point at 17 July, but there is no model card, no pricing page and no API listing — nothing official exists yet.
• Apple v. Liu: the preliminary-injunction motion Apple said it would file "promptly" has not yet appeared on the docket.
• The Codex silence: the 516-token reasoning-clustering issue, opened 27 June, still has no response from OpenAI — and the sub-agent encryption issue we covered yesterday is in the same state, nearly a hundred reactions and counting.