- Rundown
- Cosmic Rundown: Lovable Leak, ggml 1.0, Ground Stop

Cosmic AI
November 14, 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.
Many of today's links are about software that was built fast and shipped without enough checking. Research on apps made with an AI app builder found exposed user data. An AI agent deleted a professor's files. A PostgreSQL contributor asked people to stop filing AI-written bug reports. On the more positive side, ggml reached 1.0, and there were more practical guides for running large models on your own hardware. Below you'll find a security checklist, a table of agent news, notes on local inference, the FAA ground stop, and a set of quick hits.
A Security Checklist for AI-Built Apps
Lovable apps expose passport details and more than 200K private messages (discussion) is research from crissyfield.de on apps built with the Lovable platform. According to the title, the exposed data included identity documents and private messages that should never have been publicly reachable.
Before an app generated by an AI builder holds real user data, ask these questions:
- Where are access rules enforced? A check that only runs in front-end code does not protect anything.
- Which keys ship to the browser? Anything in the client bundle is public.
- Did a person review the data model? Generated schemas and storage rules often default to open.
- Who gets notified if data leaks? Decide this before launch.
Two other links make related points:
- PyPI's redesigned project pages with clearer security signals are now live (discussion). The new pages make it easier to judge a Python package before you install it.
- Christophe Pettus: Stop Using AI to Report PostgreSQL Bugs (discussion). A database maintainer asks reporters to stop passing along AI output they have not verified.
Agents: More Autonomy, Same Need for Guardrails
| Link | What it is | Discussion |
|---|---|---|
| Advancing science and engineering with Antigravity's /teamwork | Google Research on a feature that lets agents work together on technical problems | HN |
| Agweb | A browser built for agents that aims to stay out of the way | HN |
| Claude asked to renumber pages deletes professor's syllabus | An agent given a simple formatting task deleted a law professor's syllabus and a year of lecture notes | HN |
| Nonna's RAGù: AI skill files only get you started | An essay arguing that skill files are only a starting point for getting good agent output | HN |
The third row is the one to keep in mind. An agent's access should match its task. Keep anything it can touch under version control, and back it up before you let it edit files.
Local Inference Gets a Stable Base
- ggml v1.0.0 (discussion). The tensor library behind llama.cpp has reached 1.0. A 1.0 release usually means a project is ready to commit to a stable interface, which helps anyone who builds on top of it.
- Unsloth Qwen3.8-27B Quantizations, Dynamic v3.1 Update (discussion). Updated quantizations for a mid-size model that runs on a single workstation.
- Serving GLM-5.3 at 607.8 tok/s across two Nvidia DGX stations with vLLM (discussion). A walkthrough of serving one model across two machines when one GPU box is not enough.
- Show HN: DeepSeek V3 in ~200 lines of pure Jax (discussion). A compact implementation that is useful for learning how the architecture works.
- Generative AI uses more energy on non-English text (discussion). University of Michigan research that is worth reading if you run multilingual workloads.
- Nvidia allegedly engineered scarcity by funding AI startups to buy its GPUs (discussion). One more reason people are interested in running models on hardware they already own.
When One Upgrade Grounds a Country
The FAA's nationwide ground stop showed how much depends on a single piece of infrastructure:
- Show HN: Live map of U.S. flights during the FAA nationwide ground stop (discussion) showed the stop as it happened.
- FAA blames ground stop on equipment issue during telecoms network upgrade (discussion) gives the official explanation.
- Copper wire upgrade caused nationwide flight suspension (discussion) adds more detail on the cause.
According to these reports, the outage came from a routine upgrade, not an attack. Have a rollback plan ready before you change anything other systems depend on.
Privacy Regulators Stay Busy
- Italy fines Meta $887M over Android users' data access (discussion).
- Italy fines Meta over photo privacy (discussion) covers the Italian action under a different headline.
- Tech Companies Are Trying to Shift the Concept of Privacy (Again) (discussion) is the EFF on how the industry is reframing what privacy means.
Quick Hits
- Terminal borders that don't cost a cell: a clever trick for terminal UI layouts. (HN)
- Terminactory: an idle factory game for the terminal, written in Rust with ratatui. (HN)
- Let's Write a PostScript Interpreter: a hands-on language implementation tutorial. (HN)
- C3 0.8.6: library fetching, smaller safe-mode binaries, and stricter type rules. (HN)
- Formalizing US Tax Code (Using Lean): an attempt to formally specify tax rules in Lean. (HN)
- US IRS free online tax filing open source repo: the Direct File code, published by the IRS. (HN)
- Why get an expensive Pi when I have an old phone?: putting old phones to work as small servers. (HN)
- How to Create Your Own Decentralized Messenger Protocol: a guide to designing your own messaging protocol. (HN)
- A visual explanation of reinforcement learning: an interactive explainer. (HN)
- Fursoy: an out-of-order RV64 core in SystemVerilog that boots Linux. (HN)
- Reverse engineering a vintage diode-transistor logic chip: a silicon teardown from righto.com. (HN)
- Monkey Island: all 13 versions side by side: a visual history of one game across its many ports and editions. (HN)
- KDE: The documentary: a film about the desktop project. (HN)
- What happened to 3Com?: a look back at the networking company. (HN)
- A Family Tree of Software Job Titles: where today's job titles came from. (HN)
- Whiteboard: an open source alternative to Miro. (HN)
What This Means for Content Teams
Keep write access on the server. The Lovable research shows how quickly a public client can expose private data. In a headless CMS, separate read and write keys help. The Cosmic API docs say plainly that the write key should never go in a browser bundle, a mobile app, or a public environment variable. If you build with Lovable, the Cosmic Lovable guide shows how to connect content to your app.
Give agents a scoped interface. Antigravity and Agweb show agents getting more independence, and the deleted syllabus shows what can go wrong when they do. The hosted Cosmic MCP Server gives Claude, Cursor, and other MCP clients a defined way to read and manage content. Cosmic AI Agents draft and publish inside your content model.
Make mistakes recoverable. Revision history and automatic backups, listed on the features page, let you roll back content that a person or an agent got wrong.
Keep a person in the loop. The PostgreSQL post makes the same point for bug reports that applies to published content. Workflows add approval gates so an editor reviews AI-assisted work before readers see it.
To get started, read the documentation or create a free account.






