The week: a Workers spending limit, Node.js fixes 16 flaws, GitHub's brownout on Monday
สรุปสัปดาห์ 9 พฤศจิกายน: Workers ตั้งเพดานค่าใช้จ่ายได้แล้ว Node.js แก้ช่องโหว่ 16 รายการ และ GitHub ทดลองบล็อก pull_request_target รอบสองวันจันทร์นี้
Four notes: GLM-5.3 Flash's replacements measured, Supabase CLI v3 in CI, llama.cpp 1.0.0, checking AI crawlers. Plus five outside items, from a Workers spending limit to Copilot models ending 14 December.
Since 13 November a Cloudflare account can put a spending limit on Workers and Durable Objects, and when the limit is reached Cloudflare blocks new Worker invocations until the next billing period or a higher limit 1. A Worker is a small program that runs on Cloudflare's servers each time a request arrives; a loop, a bot or a viral day can run up its bill overnight. That leads the five outside items. First, the week's four notes.
Four notes, 9 to 14 November
Monday, models. GLM-5.3 Flash's replacements: Gemma 4 26B costs less per task, Qwen3.8 27B 3.6 times more: Workers AI stops serving GLM-5.3 Flash on 20 November, five days from today, and names two small models to move to. We ran ten builder tasks twice on each, 180 calls. With thinking on (the hidden reasoning a model writes before it answers) all four models got 20 of 20; Gemma 4 26B used 410 neurons, Cloudflare's billing unit, against 531 for Flash and 1,920 for Qwen3.8 27B. With thinking off the output tokens fell by 91 to 96% and the multi-step arithmetic failed. Both replacements run on the free daily allowance, which GLM-5.3 Flash does not.
Tuesday, last-mile. Supabase CLI v3 is the default since 2 November: a CI db push now stops for a yes: a deploy script that installs latest got v3 overnight. Database commands now target the copy of Supabase on the same computer unless given --remote, and a change to the hosted project fails where nobody can type yes, unless --yes is passed. We ran the same push on v2 and v3 the way a CI runner sees it. GitHub's code search found 4,728 public workflow files running supabase db push and 189 passing --yes. The note has both fixes: the v3 flags, or a pin to v2.
Friday, models. llama.cpp 1.0.0: 36 deprecated calls removed, --defrag-thold gone, FP8 KV cache on CUDA: the first major version of the program that runs open models on your own computer. Code using the 36 old function names stops compiling, and one grep in the note finds them. The KV cache (the model's memory of the conversation so far) can now be stored at 8 bits a value on NVIDIA cards, half of the default. The server refuses more by default: tools on a public address need an API key or a flag. Homebrew still gave 0.9.0 on release day.
Saturday, security. 16 of 31 AI crawler names publish an address list, one signs; reverse DNS names 3 of 12: a user agent, the name a visitor gives itself, is a claim anyone can type. We read the operator page behind each AI crawler name our traffic dashboard knows. Sixteen names have a published list of the addresses they crawl from, Exa signs its requests, six give a name and nothing to check it against. The older reverse DNS check named the operator for 3 lists of 12, because the AI companies crawl from rented cloud addresses. A 37-line checker in the note matches an address against the lists in under a millisecond.
Five items, 10 November to 14 December
Now: a spending limit for Workers and Durable Objects. In the dashboard, Billing, then Spend limits: a limit, an email and alert thresholds for each product, across the account for the current billing period 1. Turning off Block usage sends the email alerts and lets the app run on 1. The cf command-line tool sets the same thing, cf billing spend-policies create --product-id workers --threshold-cents 5000 --action usage_block for US$50 1. Cloudflare's page says reporting and blocking take time, so recorded spend can pass the limit, and it names no error code for a blocked request 1. AI Gateway's spend limits are a separate setting for Workers AI and other model bills 1. A prototype with a public URL and a card on file is the case for setting one today.
10 and 11 November: Node.js fixed 16 vulnerabilities. The releases are 26.13.1, 24.22.3 and 22.23.5 2; the release index dates the first two 10 November and 22.23.5 11 November 3. Seven of the 16 are rated high, most of them in HTTP/2 and TLS (the encryption under HTTPS), among them a use-after-free in respondWithFile() and a TLS 1.2 renegotiation that changes the server name and gets past client certificate checks 2. The high-rated ones are in Node's own HTTP/2 and TLS code, which a server runs when it handles those connections itself rather than behind a proxy. The index also lists 26.x as long-term support, code name Lithium 3. A Dockerfile pinned to node:24.22.2 or earlier needs the new number.
Monday 16 November, 16:00 to 18:00 UTC: GitHub's second pull_request_target brownout. pull_request_target runs a workflow with the base repository's secrets on a pull request from a fork 4. For those two hours, 23:00 to 01:00 in Thailand, public repositories on Free, Team and Enterprise Cloud see the runs that the 2 December default would block fail with an error annotation 4. Outside the window the rule runs in evaluate mode, with warnings 4. A failed run tomorrow night is the list of workflows to fix or to allow by name before 2 December.
8 December: Dependabot alerts pause on repositories with no activity in a year. Dependabot is GitHub's service that warns when a dependency has a known flaw. From 8 December it turns alerts off on repositories with no push, creation or alert switch-on in the past year, on every github.com plan 5. The alerts are archived, not deleted; any push turns them back on and restarts the year 5. Repositories under an enforced security configuration that turns alerts on are exempt 5. The page does not say whether Dependabot's security updates, the pull requests that fix the flaw, pause too 5. An old client site that still serves traffic from an untouched repository loses its warnings that day.
14 December: six models leave GitHub Copilot. Claude Fable 5, Claude Haiku 4.5, Gemini 3.8 Flash, Gemini 4 Argon, Grok 4.6 and Kimi K3 go from every Copilot experience; the suggested replacements are Fable 5.2, Haiku 5.5, Gemini 4 Hydrogen, Gemini 4 Carbon and Grok 4.7, with Kimi K3's to be announced 6. Haiku 4.5 stays for individual subscribers on annual plans 6. On Business and Enterprise the replacements switch on by themselves unless an administrator turned the default off 6. The page does not say what happens to a request that still names an old model after the date 6, so an integration or an agent setting that names one is the thing to change before it.
Sources
- Set spending limits for Workers and Durable Objects · Cloudflare changelog, 13 November 2026 developers.cloudflare.com
- Wednesday, November 11, 2026 Security Releases · Node.js nodejs.org
- Node.js release index (index.json) nodejs.org
- pull_request_target blocked by default December 2 · GitHub changelog github.blog
- Dependabot alerts will pause on inactive repositories · GitHub changelog github.blog
- Upcoming deprecation of selected GitHub Copilot models in December · GitHub changelog github.blog
Hongdam Studio services on this topic
- AI assistant with your name on itYour own chat assistant. It answers from your web pages, and your staff can log in.from ฿20,000Buy
- Finish my vibe codeAn app or site made with an AI tool such as Lovable, Bolt, Replit or Cursor: we read it, fix it and publish it.from ฿5,000Buy
- Your own AI website builderWe install it on your computer and accounts, and teach you to use it.from ฿25,000Buy
- Website maintenanceUpdates, backups, an alert if your site stops working, and 2 hours of fixes a month.฿3,000 / monthBuy