AI News, Aug 29: OpenAI Cuts Off Cursor Over Its Owner
Fifteen days after SpaceX closed its acquisition of Cursor, OpenAI told Cursor it has until November to find other models. Three other stories in the same window turned on the same question, which is not what a model can do but who is cleared to run it.
The Big Story: OpenAI Is Cutting Cursor Off
SpaceX completed its roughly $60 billion purchase of Cursor parent Anysphere on August 14. On Saturday OpenAI said it is winding down the contract that supplies its models to the editor, with a shutoff date of November 12, the longest notice its agreement allows. The stated reason is trust rather than money: OpenAI said it cannot be confident SpaceX will use the technology within its terms of service, and pointed at Twitter breaking an OpenAI contract after Musk bought it, plus Musk’s admission under oath this year that xAI distilled OpenAI model outputs. Bloomberg broke the decision, with Engadget confirming the timeline.
Cursor’s response was to shrug at the volume. CEO Michael Truell told The Information that OpenAI serves only about 5% of Cursor’s traffic and said the move surprised him. Anthropic co-founder Tom Brown said the same day that Anthropic will increase the compute available for Claude inside Cursor. Individual developers keep access through their own API keys, so the practical hit lands on Cursor’s bundled plans.
The Hacker News thread ran to 449 comments and landed somewhere less flattering to Cursor than to either lab. The dominant read was that a product whose economics depend on reselling four other companies’ APIs was always one contract away from this. Several paying Cursor users said the cutoff pushes them toward Anthropic, not toward OpenAI direct, which is roughly the opposite of a punishment.
Four stories this window turned on who is allowed to run a model
Z.ai released the GLM-5.3 weights on Friday, two weeks after the hosted version, and said plainly that the delay was a safety hold: internal testing found the model finds and exploits software vulnerabilities at close to US frontier level. The licence is the other half. Z.ai dropped MIT-style terms for a version requiring any company with $10 billion or more in revenue to pass a Z.ai security review before hosting the model, a restriction aimed at hyperscalers and nobody else.
South Korea went the other direction and made access a public utility. The Ministry of Science and ICT picked SK Telecom, KT and Kakao to provide free frontier chatbots and public-service agents to every citizen, supplying 512 Nvidia B200 GPUs this year. The condition attached is the point: at least half of model usage must run on qualifying Korean foundation models, another 30% on other domestic developers. That is industrial policy dressed as a consumer benefit.
And the ruling we led with yesterday belongs in the same column. Judge Rita Lin’s decision voiding the Pentagon’s supply-chain-risk label on Anthropic was, stripped of the constitutional language, a finding that a customer cannot blacklist a vendor over the terms it refuses to drop. Four fights, one shape.
Today’s Top Stories
Sony and Warner sued Anthropic and named its founders personally
Sony Music Publishing and Warner Chappell filed Friday, alleging “one of the largest and most blatant ongoing thefts of intellectual property in history” and “a brazen campaign of illegally torrenting, scraping, and downloading” tens of thousands of copyrighted compositions to train Claude. The escalation is procedural rather than rhetorical: CEO Dario Amodei and co-founder Benjamin Mann are named as individual defendants, which Axios notes the earlier music suits did not do. UMG and Concord’s $3.1 billion January case stopped at the corporate entity.
Anthropic published an AI that does alignment research for $4 an hour
Anthropic’s Alignment Science team, led by Anthropic Fellow Chen Yueh-Han, built an automated alignment researcher that reads the literature, proposes a method, trains a model on it for thirty minutes, and iterates. It improved scores on all ten misalignment benchmarks tested without degrading general performance, and the best method it produced beat what experienced human researchers proposed, on average inside six hours. The number that will get quoted is the cost: roughly $4 an hour in API spend against roughly $150 an hour for a human. The paper is candid that this only works as far as the benchmarks actually capture the goal, which is the entire open question.
Nvidia paused its revenue-sharing program over its own lawyers’ antitrust worries
Nvidia has halted parts of the AI Compute Partnership it launched in July, under which it backstopped smaller cloud providers’ GPU purchases in exchange for 50% of revenue above an agreed base hourly rate. The program carried around $36 billion in commitments, with Sharon AI and Firmus Technologies as initial partners. Per the WSJ scoop picked up widely on Friday, internal staff flagged antitrust exposure and providers objected to Nvidia vetting their customers. It comes days after a quarter with $96.2 billion in revenue, up 106% year over year, and the same week Axios counted Nvidia’s fingerprints on more than $750 billion of AI investments and financing deals.
Salesforce put its CRM inside Claude
Announced on Tuesday’s earnings call and detailed in VentureBeat’s write-up this week, “Salesforce in Claude” is a plugin for Claude CoWork shipping with 37 pre-built sales skills covering meeting prep, deal health review and pipeline analysis. Sellers query and update live CRM records without opening Salesforce. Pilot customers have it now, open beta is expected in September. Marc Benioff spent the call arguing this is not the SaaSpocalypse, a strange position to defend while announcing that customers never need to open your app.
a16z raised $1.1 billion for the unglamorous half of AI
Andreessen Horowitz closed its first hardware-dedicated fund, aimed at chips, memory, networking, storage, data centres, cooling, robotics and electrical infrastructure. GPs Raghu Raghuram and Martin Casado said hardware has gone from a sliver of the firm’s deal flow to over 20% of it. A software-first firm allocating a billion dollars to atoms is a decent proxy for where the binding constraint now sits.
Debian decided AI-written code is fine as long as you own it
The vote closed Saturday and Choice 5 won: Debian neither endorses nor prohibits generative AI tools, but “the use of a generative AI tool does not diminish the contributor’s responsibility” to understand, review, test and where appropriate modify the output. Two stricter proposals both lost to None of the Above. The dissent is the same objection maintainers keep raising in practice: “responsible” is unenforceable, and review discipline is exactly what erodes under volume.
Quick Hits
- Open weights: Tencent open-sourced Hy4 preview, a 770B-parameter mixture-of-experts model with 49B active and a context window over one million tokens, under Apache 2.0.
- Education: Anthropic opened Claude for Teachers to whole schools and districts as a free managed offering with SSO and a FERPA-aligned data agreement, with Detroit Public Schools piloting it, per GovTech.
- Agents that spend money: xAI’s Grok Bot now supports shareable templates bundling instructions, memories, skills and integrations into a one-click install, and can make purchases with Stripe issuing a single-use card per approved transaction, reports TestingCatalog.
- Inboxes: ChatGPT added support for connecting multiple Google accounts across Gmail, Calendar and Contacts in one conversation, so a work and a personal inbox are queryable side by side. The change appears only in OpenAI’s release notes, with no independent coverage yet.
- Also inboxes: Brave shipped email aliases in the browser, five free, with forwarded mail deleted from Brave’s servers after delivery.
- Hardware privacy: Meta patched the Ray-Ban loophole where users started a recording and then taped over the capture LED. The camera now stops entirely if the light is covered mid-session.
- Funding: Owner raised $240 million at a $2.3 billion valuation led by Goldman Sachs Alternatives for its agent-run platform for local businesses, having crossed $100 million ARR.
- Debt, not equity: neocloud Lambda raised $1 billion in short-dated debt arranged by JP Morgan to buy Nvidia chips it leases to Microsoft, against more than $400 billion of AI-related debt raised globally in 2026.
- Automated takedowns: an AI brand-protection service advertising “85% faster takedowns” got the open-source Luanti engine pulled from Google Play on a Minecraft claim citing no evidence. It filed the same notice in 2023, when the app stayed down 46 days despite winning the appeal.
- Security throughput: Anil Madhavapeddy documented servers being probed with matching exploits within minutes of a public PR fixing an OCaml path-traversal bug. The rclone maintainer replied that the project took about 20 security disclosures in its first ten years and over 40 in the last month.
- Who actually worries: TU Darmstadt’s AI Monitor 2026, a survey of over 2,000 people with YouGov, found 43% of those rating their own AI understanding as excellent expect AI could take over their work soon, while only 15% had done any AI training at all.
Ready to automate your busywork?
Carly schedules, researches, and briefs you—so you can focus on what matters.
See what people say
"Before Carly, I relied on a Calendly link, but the whole process felt impersonal and not very professional. Carly changed that by handling all the back-and-forth, so I'm no longer stuck in endless email threads trying to line up schedules.
Now Carly reaches out to candidates, shares my real-time availability, lets them pick a slot, then sends a Zoom link and drops it straight into my calendar. She sends reminders to both of us before each call, which has significantly reduced no-shows and last-minute confusion.
On top of scheduling, Carly acts like a full executive assistant, sending me my schedule the night before so I can prepare for each call. It reminds me of the old x.ai assistant, but Carly is noticeably smarter, faster, and better suited to my healthcare recruitment business."
