- Rundown
- Cosmic Rundown: GPT-6.2, Optimizely Hack, Alloy CSS

Cosmic AI
November 17, 2026
This article is part of our ongoing series exploring the latest developments in technology, designed to educate and inform developers, content teams, and technical leaders about trends shaping our industry.
Today's links cover a supply chain attack that reached some of the largest sites on the web, two new OpenAI models, and a series of outages across Apple, a US government network, and a brokerage. The post also covers a CSS library that works with JavaScript turned off, a new Chrome feature for rendering HTML inside a canvas, some database reading, and quick hits.
The Optimizely Compromise: A Checklist for Third-Party Scripts
Two reports describe the same incident. Optimizely compromised, served malware on Zoom, GitLab (discussion) comes from Report URI. Optimizely hack infected CNN, Fox News, BBC, Salesforce sites with malware (discussion) comes from BleepingComputer. The Times reports that 178 websites were hacked (discussion), including BBC, Zoom, and IKEA.
The coverage is linked above. This section focuses on what to check if your site loads JavaScript from another company:
- List every script that comes from another domain. That includes tag managers, A/B testing tools, analytics, and chat widgets. Many teams cannot produce this list quickly.
- Set a Content Security Policy. A strict
script-srclimits where code can load from. Start in report-only mode to see what your pages actually load, then enforce the policy. - Use Subresource Integrity for static files. SRI pins a file's hash, so the browser will not run the file if it changes. It does not work for scripts that a vendor updates in place, and those are the scripts to review most carefully.
- Decide where each script runs. An experimentation tool on login and checkout pages carries more risk than the same tool on a marketing page.
GPT-6.2 and the Harness Question
OpenAI introduced GPT-6.2 Astra and Sol (discussion). A separate submission notes that OpenRouter lists GPT-6.2 Astra at $8 per million input tokens and $40 per million output tokens (discussion).
Another thread is useful for teams choosing tools. What is the best harness for each model? (discussion) introduces Runta's FrontierHarness eval, which compares how models perform in different agent harnesses. On a launch day, remember that a benchmark score measures a model together with the harness around it.
Related reading:
- How the Claude Code team thinks about verification (discussion): how one team checks agent output before trusting it.
- PINE: A Reactive Harness for Agents (discussion): a new open-source harness project.
- TrnRT, an open-source LLM inference engine for AWS Trainium (discussion): inference on hardware other than Nvidia GPUs.
- Gemini 4 Hydrogen: a faster model for general purpose robotics (discussion): Google's new model for robotics.
A Rough Day for Uptime
Several outages reached the front page within a few hours of each other:
| Outage | Source | Discussion |
|---|---|---|
| Apple services down | Downdetector | HN |
| Apple iCloud account and sign-in outage | Apple System Status | HN |
| Highly classified US national security communications network goes down | CNN | HN |
| Same network outage | Reuters | HN |
| Charles Schwab website goes down | Business Insider | HN |
| FAA pauses Juniper router installations after nationwide ground stop | WSJ | HN |
The Apple sign-in outage is a reminder that a shared identity provider failing can take down everything that depends on it. List which of your systems rely on a single sign-in, DNS, or network vendor. Host your status page somewhere that does not share those dependencies.
Web Platform Notes
- Alloy CSS: dialogs, menus, and tabs that work with JavaScript off (discussion): a Show HN that builds common interactive components on HTML and CSS, with no script required. It is a good example of progressive enhancement.
- HTML-in-Canvas in Chrome 157 (discussion): interactive DOM inside 2D, WebGL, and WebGPU canvases. This matters for canvas-heavy apps that have struggled with text rendering, forms, and accessibility.
A Short Data Reading List
- Rows vs. Columns: How different are they, really? (discussion) from SpiralDB, on the line between row and columnar storage.
- VennDB (discussion): an append-only, in-memory Rust database where rows are queried using bit flags.
- Replacing implicit optionals with a None type in TypeDB (discussion): a language design change in a database query language.
- LanceDB raises $57.5M (discussion): funding for multimodal data infrastructure built for AI workloads.
Compilers, Old and New
- Nimony: A New Compiler for Nim (discussion)
- Turbo Pascal Compiler Written in Turbo Pascal (discussion)
- VisualBasic.World (discussion): a gallery of Planet Source Code programs from 1997 to 2012.
Quick Hits
- Carnegie Mellon is putting computer science courses on YouTube for free (HN)
- BugMeNot has closed: the shared-login service shuts down. (HN)
- We don't need more apps. We need better open source: an essay on where effort should go. (HN)
- Antfly learned to stop worrying and love Apache 2.0: a licensing change explained. (HN)
- If an app asks for Full Disk Access or System and App Protection Override: Apple's guidance on high-risk permission prompts. (HN)
- Exploiting Replay Protected Monotonic Counters to Brute-Force Windows Hello PINs: new research on PIN protections. (HN)
- Why 3dfx cards had washed out color output: John Carmack on a classic GPU quirk. (HN)
- Chemistry is Actually Geometry (HN)
- Comcast's "fiber-powered Wi-Fi" is cable, AT&T says: a false advertising case. (HN)
- New evidence shows school phone bans are gradually improving test scores (HN)
What This Means for Content Teams
Know what runs on your pages. The Optimizely incident shows how much trust a single script tag carries. With a headless CMS, content is delivered as data through a REST API, and your front-end team decides which scripts ship with it. That makes the third-party script list shorter and easier to audit.
Test models on your own work. The harness eval shows that one benchmark score does not tell you how a model will perform in your setup. Cosmic AI Agents run inside your content model, and Cosmic Workflows add approval gates, so an editor checks AI-assisted drafts before they publish. That follows the same approach as the Claude Code team's post on verification.
Give agents a narrow interface. The hosted Cosmic MCP Server exposes a fixed set of content tools to Claude, Cursor, and other MCP clients. Separate read and write keys keep write access on the server.
Build for partial failure. The day's outages and Alloy CSS lead to the same idea: pages should still work when something upstream breaks. Content served from an edge-cached API and rendered on the server stays readable even if a client-side script fails. The features page covers the CDN, revision history, and backups.
To get started, read the documentation or create a free account.






