HU
Idea log · Friday, November 13, 2026

Critical Apache HTTP/2 flaw under WordPress sites

Based on 15 newsletters · 8 to try · 4 to watch · 5 in brief
Today's most important item is K1: if your WordPress site runs on Apache with enabled, a critical flaw may affect it. A plugin update won't fix it; the hosting provider or a server update will.

🛠 Ideas to try

Critical Apache flaw: worth asking your hosting provider

urgent

A critical flaw () in the module of the Apache web server was fixed on Nov 11: it allows remote code without logging in. Affected versions are 2.4.55–2.4.70; the fix is in 2.4.71. A server is vulnerable only if Apache itself offers connections; if a or nginx sits in front of it, traffic to Apache is often plain HTTP/1.1. Updating WordPress, plugins or Wordfence won't fix this, because the flaw is in the layer below PHP. Distributions often the fix, so the version number alone doesn't tell you whether a server is affected. The flaw was first reported by Anthropic's OSS Scanner (Oct 26). Public crash-inducing test code already exists; no attacks in the wild are known yet. According to a web search, this is already the second critical mod_http2 flaw of 2026 (the first was CVE-2026-23918 in May).

Source: Wordfence · wordfence.com/blog/2026/11/cve-2026-119618-apache-http2/
For you: If you run a WordPress site on shared or managed hosting (e.g. yourdomain.com), send your provider this question: "Does the Apache server hosting my site accept h2 or h2c connections, has the CVE-2026-119618 fix been installed, and was Apache restarted afterwards?" Until it's fixed, ask them to temporarily disable .
For you: If you run the server yourself, update to 2.4.71, or remove h2/h2c from the Protocols line and restart Apache. Make a fresh backup of the files and the database first.

Gemini Decisions : cheap labeling for high-volume text

Google's new model (Gemini Decision 1) doesn't write text; it gives probabilities for predefined questions: yes/no, multiple choice (2–255 options) or rubric scoring. In one request it evaluates up to 1,000 questions on the same input (text, image, PDF, audio, video), in 150–250 ms. Every answer comes with a confidence score, so uncertain cases can be passed to a stronger model. ❓ The sources conflict on price: Google's blog says $0.10 per 1M input , while TLDR wrote "$0.10 per 2M tokens". Output is free. The page says nothing about EU availability.

Sources: TLDR, TLDR Dev, TLDR AI, The Code, The Neuron · blog.google/innovation-and-ai/technology/developers-tools/gemini-decis
For you: If you want to sort newsletters or emails with AI, pre-filtering (promo / worth watching / questionable / discard) can be done with this, with a single "choice" question per email, and only the relevant emails go to a more expensive model for summarizing. That saves most of the cost.

, or skill? How many tokens tool descriptions cost

OpenRouter measured that the same GitHub task on Sonnet 5.5 cost 1,421 tokens with the gh command-line tool (), 3,315 with a , and 59,417 with the GitHub server's 46 default tools. The reason: every tool's description is sent to the model again with every request. (deferred loading of tools) roughly halved the cost on Claude, with mixed results on other models. Their recommendation: use a where one exists; with , load only the tools you need, and trim large tool responses (requesting only the needed fields).

Source: TLDR AI · openrouter.ai/blog/insights/mcp-vs-cli-vs-skills/
For you: If you're building your own AI app, don't hand it the full Gmail or other as a tool. Let a small function fetch the data and send only the cleaned text (without HTML and link lists) to the model; that cuts the token count to a fraction. In Claude Code, it's worth checking now and then which servers are enabled for no reason.

Anthropic's official Fable 5.2 prompting guide

  • start at high, then check whether medium gives good enough results. Effort sets how much thinking happens, not how long the reply is.
  • Against stopping early the guide includes a "Finishing work" block ("don't stop while you still owe the work you were asked for…"). In Anthropic's tests it cut abandoned sessions roughly in half.
  • Approvals specific approval works better than a bare "continue" ("Yes, run the migration on the test database"). Word prohibitions precisely too: "ask for approval before deleting files" works better than "never delete files".
  • Clean up old rules delete outdated lines like "never retry", "think less" or "don't delegate", and artificial token-budget text. For , check which model runs by default, since they can account for much of the cost. Also review the instructions coming from the .
Source: The Code · platform.claude.com/docs/en/build-with-claude/prompt-engineering/promp
For you: If you use Claude Code or your own system prompt, review your CLAUDE.md and prompts against K4d. The K4b block fits well into scheduled, unattended runs.

"Check the translation, not just the answer" – a prompt for spreadsheet analysis

AI often answers a slightly different question than the one asked: it calculates net instead of gross, or counts only active employees instead of all of them. The Neuron's method: (1) have the AI list every definition and assumption it used; (2) compare that with the wording of the request; (3) in a new chat, give only the request and this list, with this prompt: "Below are my original request and the definitions and assumptions an assistant used. List where a definition is narrower, broader or different from my wording, and for each, describe how the result would change with my original wording."

Source: The Neuron
For you: If you keep timesheets in Excel and AI builds formulas or summaries for them, ask about these typical slips: "hours worked" with or without leave, weekend premiums, and how part-timers are handled.

GitHub security: 4,837 hijacked accounts and locking GitHub Actions

  • Check using stolen passwords, attackers pushed credential-stealing code from 4,837 GitHub accounts into repositories, hidden in workflows or an npm postinstall hook (which also runs on the developer's machine during npm install). Most of these changes were quickly reverted and the logs deleted. Telltale signs: a wfx-docs, ci-check-NNNN or chore/xxxxxxxx branch; a titled "docs: dependency audit notes"; a "ci: refresh build cache" or "chore: add ci helper" commit; a .ci-xxxxxx.sh file. If you find any: revoke tokens and SSH keys, turn on two-factor authentication, and rotate every secret the repository could access. Sources: The Code · thehackernews.com/2026/11/4837-hijacked-github-accounts-ran-ci.html
  • A for gh extension install github/gh-actions-lock, then gh actions-lock in the repository root. This creates .github/workflows/actions.lock, which records exactly which commit each version tag points to. In , gh actions-lock --verify checks it; update it automatically. Source: TLDR Dev · hugovk.dev/blog/2026/lock-github-actions/
For you: If you have code on GitHub, you can check for the signs above in five minutes (especially commits after Nov 10). If you deploy your website with GitHub Actions, K6b is worth doing too.

Removing the "AI smell" from articles

A list of typical LLM phrasings is going around. Examples: openings of the "not X, but Y" kind, announcing "this is crucial" instead of showing what the thing does, punchy closing lines, lists of three out of habit, and "happy to elaborate" at the end. The list comes from the "declauding" skill in oaustegard/claude-skills: the skill rewrites the draft, and in a separate mode marks in HTML what it changed and why. In Claude.ai it can be uploaded as a ZIP under Settings / Skills.

Sources: Superhuman · github.com/oaustegard/claude-skills/releases/tag/declauding-v0.13.0
For you: If you write local news, sports articles or blog posts from AI drafts, this skill (or the list pasted into your prompt) gives more natural text. Most of these phrasings show up in other languages too.

Code quality add-ons for Claude Code

  • Modern Web Guidance 1.0 Chrome's skill pack (with Firefox and Edge contributing) gives coding agents best practices for accessibility, performance and security, and includes compatibility data. Install: ; also available in the Claude Code plugin marketplace. Source: TLDR Dev · developer.chrome.com/blog/modern-web-guidance-v1
  • A one-line bug-hunting prompt Find all of the implicit state machines in my app. It finds state handling scattered across the code (), which produces "impossible" states and hidden bugs. One developer found 212 in the Plane codebase. Source: The Code
  • shadcn/improve a skill that reviews a codebase and writes a prioritized improvement plan that an agent can then carry out. Source: TLDR Dev · github.com/shadcn/improve
For you: If you're building a new website with an agent, K8a gives you accessible, fast HTML/CSS from the start. Run K8b and K8c once on your existing Claude Code projects.

👀 Worth watching / ❓ Raises questions

A new generation of AI scams

Scammers use social media posts to write personalized phishing emails (). In one case, a "CEO" and "lawyer" were used to extract $129M. Even small organizations should give staff a short warning: always verify urgent payment requests through a second, known channel. (Superhuman)

Nod – n8n's personal agent that runs on your own machine

A free, open-source (Apache 2.0) research preview for Mac, Windows and servers. It tracks invoices, drafts replies, watches pages, and asks for approval before sending. It also works with local models (Ollama), but those need tool calling and a 64K context. The question: how far can you trust an experimental tool with access to your email and calendar? It could run on an always-on home machine. (Superhuman, The Neuron · nod.n8n.io/)

Leaked keys → a $740,000 AI bill

Intruders at the Railway cloud host stole , mainly AI provider keys, and used them to run up charges. Grok Build's 35,000 public apps also exposed confidential data. Lesson: put a monthly spending cap on AI provider keys, and keep them out of public repositories and code. (TLDR InfoSec)

"Lazarus" – a backdoor by editing fewer than 10 neurons

Microsoft researchers showed that by rewriting a few rarely used neurons in an model, a hidden command can be planted that activates on a trigger phrase. When running a model locally, download it only from official sources. (TLDR InfoSec, TLDR IT)

📰 In brief

Jensen Huang has stepped down as head of Nvidia; CFO Colette Kress is leading the company on an interim basis. The sources differ on the details: TechCrunch reports a resignation, others an "ouster", and reports describe a second, or per Reuters a third, affair. (TechCrunch ×2, TLDR, TLDR AI, Superhuman)
Anthropic introduced Claude Meetings (beta, Team/Enterprise), Sheets and Whiteboards, plus interactive connectors; Claude Dashboards became generally available as Claude Data. (TLDR IT, TLDR Data, The Neuron, Claude Code Daily)
AWS Emulator runs 200+ AWS services locally, without an account, and was written entirely by AI agents. (TLDR, TLDR Dev, The Code, The Neuron)
Amazon Ring unveiled a $250 indoor drone camera and a $1,000 patrolling robot; the drone flies only when an alarm trips or on a scheduled round. (Rundown Robotics)
United Airlines is blocking unauthorized AI agents from booking, Amazon is restricting them too, while Walmart lets them in with permissions. (The Neuron, TLDR AI)

Based on 15 newsletters, summarized with the help of AI. Items are short summaries in our own words; follow the links for details.

🔎 Search and filter across all days