September 21: 18-0 at StarCraft and still a beginner, Amazon cuts off Meta's Muse, and Zhipu opens ZCode

Monday's scan across 72 of the 86 configured seed accounts returned 1,225 posts. Of those, 97 were published in the 24 hours of Monday, September 21, 2026, Beijing time, 34 were retweets and are left out here, and nine original posts cleared 100 likes. The seed accounts cover roughly 2.2% of @hwwaanng's network of about 3,850 accounts, and fourteen of them returned an upstream error this morning, so read this as a sample of a large network rather than a complete scan of it.

At 03:02, Peter Steinberger (@steipete), who builds the OpenClaw agent, posted one line — "New benchmark dropped" — over a link to a report at bw.swerdlow.dev. The post drew 1,631 likes, 163 reposts and 116 replies, more than anything else published in the window. 1

What the link opens is Brood War Bench: 171 matches of StarCraft: Brood War in which every player is an AI agent. Nineteen configurations each play every other one once, on rented Freestyle machines, and the harness saves the game engine's data alongside both agents' logs for each match. A configuration is one model run at one effort setting, so the "xhigh", "medium" and "low" rows below are the same model given more or less time to think. APM counts the commands an agent issued per minute; the cost per game is a token-based estimate for the Codex and Claude rows. 2

Wins and losses do not always total 18 because some matches ran to the benchmark's time limit and were recorded as neither. All figures in the table come from the report. 2

Four findings run through the write-up. Codex's strongest idea was disruption: in Protoss games it often sent a single Probe across the map to attack workers or buildings, and opposing agents spent dozens of seconds working out what to do about one Probe instead of doing anything else. Those same configurations were weak at sustained production — they delayed technology, trickled one or two basic units into defended bases, and threw workers into last stands. Astra also split itself into subagents for the economy, army production and army control, and the subagents did not talk to each other, so each new unit went straight into an attack without waiting for the rest. Grok 4.6 spent its games between actions: in one match the highest-effort Grok logged 11,138 reasoning tokens and issued six command batches across 43 minutes, and never fielded a fighting unit. Claude Fable read as the one model that wanted to play the game — it built an economy and climbed the technology tree rather than stopping at the first unit available, reaching a Lair, a Spire and Mutalisks in one win, and a Robotics Facility, Citadel of Adun, Observatory and Templar Archives in another. 2

The report's own summary of the whole tournament is that none of the models played beyond a beginner level, and that a beginner playing a photon rush would win every one of these games. The benchmark began as a version of Brood War that could only be played through agents, built to play with friends. 2

At 22:35, Steinberger posted "The beauty of running a claw yourself: they cannot block you," over a post from BuccoCapital Bloke (@buccocapital) reporting that Amazon had cut off Muse, Meta's personal AI agent. Steinberger's post drew 973 likes, 44 reposts and 99 replies; the post underneath it had 5,192 likes and 2.1 million views. 34

Muse users began seeing a popup on Sunday night telling them that "continued access by an unauthorized AI agent violates Amazon's Conditions of Use, to which our customers have agreed." GeekWire reported it first, and Amazon's account is that Meta never told it Muse would reach into the store, that the agent does not identify itself while browsing, and that it appears to capture and store customer credentials. Amazon also says Muse can open account pages and order history when a customer asks it to, and that it asked Meta to take Amazon out of the experience. 5 The Verge, which had bought tank tops through Muse while testing the agent, reports that Meta's public position is that Muse has no visibility into people's passwords or payment methods. 6 Meta's own launch material describes the design underneath the dispute: where a service offers a public API, Muse connects to it with credentials the user provides, and where a service offers none, the agent works it through a browser the way a person would. 7

Muse runs on iOS, Android, muse.ai and WhatsApp, and a week after its September 8 launch it was the top free app in the US App Store, ahead of ChatGPT. 5 Amazon's record of keeping other companies' agents off the storefront is a year long: it sued Perplexity over the Comet browser, won a preliminary injunction in March, lost it on August 4 when the Ninth Circuit held that the user, rather than the AI company, was the one accessing Amazon's computers, and was denied a rehearing on September 10. It has also blocked shopping agents from Google and OpenAI. Its own two agents, Alexa for Shopping and Buy for Me, are the exception; Amazon says Buy for Me identifies itself and lets retailers opt out. The money behind the position is advertising — Amazon took more than $68 billion in ad revenue last year, most of it from people browsing its pages — and The Register tested the other side of the block: asked to find and cart an ergonomic chair, Muse came back reporting an anti-bot wall that stopped its browser before the search page, and said it would not push past it because the notice said continuing would violate the terms. 8

At 01:36, Omar Shahine (@OmarShahine), a Microsoft corporate vice president who describes himself as working on OpenClaw for Microsoft 365, announced that the agent can now take a call: "You can call your claw, your claw can call you." The feature carries audio for now, he wrote, and it only crossed the finish line with Opus 5 and GPT-Sol 5.6; it supports GPT-live, and it is experimental and due in the next release. His post drew 297 likes. 9 At 03:02, Steinberger picked the feature up with "Your claw can now FaceTime you!", which drew 851 likes and 42 reposts. 10

At 22:57, Steinberger posted a second line, this time over a post from josh avant, who works on OpenClaw: "Proud to say that we did our homework. They found nothing critical." His post drew 365 likes and 14 reposts; the post he lay over drew 82. 1112

The audit is with Trail of Bits, run through OpenAI's Patch the Planet initiative, and OpenClaw's own write-up went up at 22:00 the same evening with the counts: 27 private repository advisories and three standalone hardening pull requests. 131415 Twenty-four of the advisories came with a severity rating and 23 were confirmed vulnerabilities, one of them fixed before it was reported; the ratings were 0 Critical, 2 High, 16 Medium and 6 Low, so "nothing critical" holds at the top rating while two findings ranked High. The remaining three were defence-in-depth findings with no rating because they did not cross a documented trust boundary. Every actionable issue was repaired and all three hardening pull requests were merged, shipped in the 2026.8.1 and 2026.7.33 LTS releases. 13

Most of the findings shared one cause: permission that did not travel with the request. Work that began with limited access could start follow-on work without those limits, and a setting turned off in configuration did not always reach a run already in progress or a permission that had been cached. OpenClaw says its tools now check the current setting each time they act, and that an approval is bound to the exact file, identity or action that will be used. 13

ZCode is Zhipu's AI coding tool, and in mid-September its users found that, while they were signed in, it packaged the whole workspace — source files, git history, deleted configuration files and keys — and uploaded the archive to an Alibaba Cloud bucket. Zhipu traced the uploads to a code-repository-index feature that shipped switched on, apologised on September 18, said the data existed only to build a wiki page in the cloud and was destroyed afterwards, and promised to open the source. 16 A report on Sina Finance describes what came next from a paying enterprise customer, 承明科技: an outside forensic examination, then a letter to Zhipu saying the uploads fired automatically and in bulk, that what moved was the full source tree and staff information rather than code snippets, that an upload was still detected on the morning of Zhipu's apology, and that the client's network requests pointed at a Singapore entity while its contract was signed with the Beijing company. The letter asks for a written answer by October 10 and lists twelve demands, among them proof of deletion and the path the data took; Zhipu had not answered it publicly when the report ran. 17

On Monday Zhipu published the ZCode codebase on GitHub and said version 3.14.0 removed the wiki entry and cut the path that generated and uploaded the local snapshot. 18 Two outside firms reviewed the result: the China Academy of Information and Communications Technology found the production bucket holding no data in the cloud, and NSFOCUS confirmed the bucket and every object in it had been deleted, with no path left that could trigger another snapshot or send files out. Zhipu also promised that no code was retained on its servers and that none of it was used to train models. 16

At 13:30, Nyarime (奶昔🥤, @realNyarime), who runs a networking business and writes about eSIMs, replied to that criticism with a defence of Zhipu's position. His post quotes a user who had called z.ai a credit to Chinese AI and now called it 一坨, a lump, and answers that Zhipu getting ICP filing — the registration a service needs before it can be served from inside mainland China — for its own AI domain is a privilege in itself: look at where the company sits, he wrote, beside the education-network operator CERNET with Tsinghua behind it. His post drew 116 likes and 23 replies. 1920

At 14:18, 郭宇 (@turingou), a developer who has retired and lives in Tokyo, quoted BytePlus's launch of Dramagic: an enterprise platform for making short dramas that, in BytePlus's description, carries a script through analysis, asset setup, storyboard and video preview, with consistency checks built in and multi-user collaboration for review. 21 The launch post drew 1,082 likes and 142 reposts. Guoyu's reaction asks why ByteDance would build the tool itself as a separate enterprise product instead of folding it into CapCut; it drew 233 likes, 21 reposts and 224 bookmarks. 22

The account that announced Dramagic is BytePlus Global, which describes itself as ByteDance's enterprise AI and cloud arm. 21

At 00:46, BunnyLau (@BunnyxStudio), a product designer and SwiftUI developer, answered the obvious question about the iPhone Duo layout study he had posted the previous evening. That study showed three arrangements of Safari's navigation on a phone that opens into a wider screen, and found two of the three unshippable: the moment a page uses sticky navigation or a floating bar, elements are covered and stop responding to taps. He also argued that Apple and WebKit could handle those viewport cases at the rendering layer. 23

His follow-up takes the next question — why not simply move the page's navigation bar to the top — and reports the test. On minimal pages with clean markup it looks slick; in the wild, odd styling and sticky overlays break it instantly, and the layout comes out more fragile than it is worth. Expecting millions of websites to adapt by hand for a foldable form factor is, he writes, a non-starter. The follow-up drew 147 likes and 16 reposts; the study it answers drew 118. 24

iPhone Duo is Apple's first foldable, introduced on September 9: a 7.6-inch inner display and a 5.4-inch outer one, titanium, with pre-orders opening October 16 and availability on October 23. 25

At 22:40, Jacob Titus (@jacob__titus) quoted Joseph Alessio (@alessio_joseph), the brand lead at Cognition, who had posted "something about clarice lispector book covers >>>" over an image that morning, and added "That Penguin can cook" with four images of his own. "Can cook" is the internet's shorthand for doing something well, and the publisher being praised here is Penguin, which keeps the Brazilian novelist Clarice Lispector in print. 262728 Titus's post drew 410 likes, 27 reposts and 119 bookmarks; the post he quoted drew 59.

At 12:14, Nyarime posted a short clip captioned 我很开心, "I'm very happy", quoting an identical post from CuiMao (@CuiMao), an AI developer in Los Angeles, from nine minutes earlier. Neither caption says what the clip shows, and the replies are where its subject appears: one tags @JensenHuang, one jokes that the Nvidia chief executive's leather jacket is missing, one asks whether this is what tokens are for, and Nyarime answers another with 那确实是苏妈 — "that is indeed Su-ma", Lisa Su of AMD. 2930 Nyarime's copy drew 723 likes, 48 reposts and 58 replies; CuiMao's drew 459 likes and 306,000 views.