✦ AI News for Builders
Nobody shipped a model today. Four things changed underneath you anyway
Monday, August 31, 2026·8 min read·4 stories
No launches this morning, which makes it a good day to look at the things that move without an announcement. Anthropic spent the weekend signing people out of Claude because ordinary malware on their own laptops stole live sessions — no password required. OpenAI is pulling its models out of Cursor on November 12 for reasons that have nothing to do with Cursor. Claude Code's weekly allowance drops 17% below today's level on September 14, which the first version of the announcement did not say. And McKinsey found that roughly a third of organisations have already skipped a software purchase because they could build the thing themselves. A session cookie, a dropdown, a quota, and a market. None of them are code you control.
No sponsored or affiliate links in this digest — the links below are sources only.
Story I
Your Claude session was stolen without anyone touching your password
Anthropic began contacting affected users on August 30 after discovering that commodity infostealer malware, running on people's own computers, had lifted their active Claude login sessions. Not passwords. Not two-factor codes. The session cookie itself — the token that says you already proved who you are. Attackers loaded those cookies and burned through paid usage limits on accounts that never saw a failed login. Anthropic revoked the compromised sessions, stripped saved payment methods from affected accounts, and refunded the unauthorised charges it identified. The families named are the usual commodity set: Vidar, LummaC2, StealC, RedLine and Acreed on Windows, with Atomic Stealer on a smaller number of Macs.
Anthropic was precise about one thing and it is worth repeating in full, because the alternative reading will spread anyway: "We have no reason to believe that this malware is related to Claude, installed through Claude, or related to anything you did with Claude." That is almost certainly true. These stealers do not target anybody in particular. They vacuum a browser profile and sell whatever is in it, and a Claude subscription is now simply one of the resellable things in there.
The sentence that should actually change your afternoon is the other one: "Signing you out of Claude stops the stolen sessions, but it doesn't remove the malware." A stealer that had your Claude cookie had everything else in the same profile at the same moment. The Claude charge is not the incident. It is the receipt for an incident you have not scoped yet.
For builders
Do not stop at rotating your Claude password, because the password was never the thing that leaked. Modern stealers take files as well as cookies, so check the ones on a developer machine that are worth the most: ~/.aws/credentials, ~/.npmrc, ~/.pypirc, ~/.config/gh/hosts.yml, ~/.docker/config.json, ~/.ssh/, and every .env in every project folder you have cloned this year. Anything with a long-lived token in it is now public until you rotate it. Then revoke sessions rather than just changing passwords — GitHub, your cloud console, npm, your registry, Vercel or Fly, and your billing dashboards all have a "sign out everywhere" control, and it is the only action that invalidates a stolen cookie. After that, the structural change worth making while the fright is fresh: publishing and deploy credentials should not live on the same machine you browse the web with. Move them to CI with short-lived OIDC tokens so there is no static secret on disk to steal. Finally, look at your billing history rather than waiting to be told — Anthropic caught this campaign, but an attacker quietly using your key at a normal rate looks exactly like you having a productive week.
Story II
OpenAI is pulling its models out of Cursor on November 12, and none of the reasons are technical
OpenAI announced on August 29 that it will end Cursor's access to its models on November 12, 2026. The cause is corporate paperwork. SpaceX closed a reported $60 billion all-stock acquisition of Anysphere, Cursor's parent company, on August 14, and OpenAI's agreement carried a change-of-control clause with a limited window to cancel. OpenAI used the window with about two weeks to spare. Its stated reasoning, in its own words, is that it "cannot be confident that SpaceX will use our technology within our terms of service, based on our experience with Elon Musk's companies violating contracts."
Scope matters here, and the headlines mostly blur it. Cursor is not shutting down and is not losing its other providers — Anthropic, Google and xAI models stay. What disappears on November 12 is a set of rows in your model picker. If your work in Cursor already runs on Claude or Gemini, this is a news item. If you have a tuned workflow, a custom mode, or a team config that names an OpenAI model, it is a dated migration with 76 days on the clock.
The transferable lesson is not about either company. It is that a capability you depend on left because a holding company changed hands, and no amount of engineering discipline on your side would have predicted it or prevented it. Model availability is a contract term wearing an API's clothes.
For builders
Spend twenty minutes finding out whether this is your problem before you decide it isn't. Grep your repo for pinned model names — grep -rn "gpt-\|o3\|o4-mini" .cursor/ .vscode/ docs/ scripts/ — and check your Cursor rules files, custom modes, and any shared team settings that specify a model rather than inheriting the default. Then do the migration while it is cheap: pick your three most repetitive real tasks, run them for a week on a model that will still be there in December, and note where the prompts break. Prompts that were tuned against one model's habits usually need rewriting rather than re-pointing, and you would much rather discover that in September than in the second week of November. The bigger habit is worth adopting today too: if a model is load-bearing for your product, route it through a thin abstraction with a tested fallback, and treat "which provider" as a config value rather than an assumption. And if you resell or build on someone else's model under contract, this is a reasonable morning to read your own agreement for change-of-control and termination-for-convenience language, because you now have a concrete example of what those clauses do when someone actually pulls them.
Story III
Claude Code's weekly limit goes up 25% and down 17%, and both numbers are true
Anthropic is making a 25% increase to Claude Code's standard weekly limits permanent across Pro, Max, Team and seat-based Enterprise plans. It is also ending the temporary 50% boost that has been running all summer, on September 13. Take a baseline of 100: you are on 150 today, and from September 14 you will be on 125. Both the increase and the cut are real, they are just measured from different starting points.
The first version of the announcement led with the 25% and did not spell out the other half. Developers added a Community Note to the post, and Anthropic deleted it and reposted with the missing sentence stated plainly: "Compared to today, this works out to a 17% reduction in weekly limits on Claude Code." Credit where it is due — the correction was quick, unambiguous, and did not hide behind a footnote. It should not have taken a Community Note.
One detail from Anthropic's own framing is more useful than the percentages: weekly allowances do not translate into a fixed number of prompts, because consumption varies with conversation length, model choice, tool use and effort level. That cuts both ways. It means you cannot compute your September headroom from a prompt count. It also means the 17% is recoverable through configuration rather than through spending, if you know where your tokens actually go.
For builders
Measure now, while the boost is still live, so you have a real number rather than a feeling. Watch your consumption for a full working week, then multiply by 0.83 — that is your September reality, and if it lands above your limit you have thirteen days to fix it rather than a surprise on a Monday. The levers that actually move the number, in rough order of payoff: drop the effort or thinking level for mechanical work like renames, test scaffolding and boilerplate, since most of a session's tokens go to reasoning you did not need; start a fresh session per task instead of letting one conversation accumulate a codebase's worth of context, because every turn re-sends the whole history; and stop re-reading whole files when a targeted search would do. If you run a team on seat-based plans, do this before September 14 rather than after, because the failure mode is not a bill — it is three engineers hitting a wall at the same hour on a Thursday, which is a far more expensive surprise than a price increase would have been.
Story IV
A third of companies skipped a software purchase because they could just build it
McKinsey's State of AI 2026 survey, covering 1,719 professionals and business leaders worldwide, found that nearly a third of organisations have declined to buy a software product or feature because they could build the functionality in-house with agentic coding tools. The sector split is the interesting part: technology highest at 41%, healthcare payers and providers at 39%, professional services and energy and materials at 38%, financial institutions at 36%, media and telecom at 34%, pharmaceuticals and medical products at 33%. That is not a tech-industry quirk. It is close to uniform.
The same survey is much less flattering elsewhere. Only 37% of respondents report any EBIT impact from AI at all, flat against 2025, and just 6% attribute 5% or more of earnings to it. Among organisations above $1 billion in revenue, the share scaling agents in at least one function went from 27% to 40%; smaller firms stayed flat at 22%. McKinsey's own summary of the tension is fair: conviction is growing faster than the returns anyone can point to.
Read those two findings together and the story is not that AI is generating value. It is that value is moving. Money that used to leave as a software invoice is now staying inside as an internal build, and it has not yet shown up as profit on either side of that trade. If you sell software to businesses, that is a churn signal arriving before it arrives.
For builders
If you sell software, run the clone test on your own product this week and be honest about the answer: could a competent team with a coding agent rebuild your core loop in a week? For a surprising number of small B2B tools the answer is now yes, and if it is yes for you, your moat was never the code. It is one of four things — proprietary data, integrations that took years of partner agreements, compliance or certification your buyer cannot self-attest, or a domain where being wrong is expensive enough that nobody wants to own the build. Identify which one you actually have, write it in a sentence, and point next quarter's roadmap at it instead of at features a weekend of prompting could reproduce. If you cannot name one, that is the finding, and it is better to have it in August than at renewal season. On the buying side the same test runs in reverse: before you renew a seat-based tool, cost the build properly — not the prototype, which agents make look trivial, but the five years of maintenance, security review, on-call and edge cases that follow it. The organisations in that 32% who priced only the prototype will be the ones quietly re-purchasing in 2027.
— The Vibe Gate news desk. We read the firehose so you can keep building.