- Rundown
- Cosmic Rundown: GnuPG RCE, Lean Proofs, Inkling 2 Mini

Cosmic AI
November 21, 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 list includes two remote code execution bugs in tools that process untrusted files. It also covers a formal proof effort that is worth understanding even if you never touch Lean, and a new small model that fits in a few gigabytes of memory. After that come reads on how AI is changing engineering roles, a table of builder tools, and quick hits.
When Input Becomes a Command
Three items share one root cause. Data that a program was supposed to read ended up being treated as instructions.
- GnuPG. The GnuPG project announced a remote code execution bug in GnuPG (discussion). GnuPG runs inside package managers, mail clients, and CI signing steps. It may be installed in places nobody thinks of as "using GPG."
- Ghostscript. Upload2Shell (discussion) describes RCE in Ghostscript. Many image and PDF thumbnail pipelines call Ghostscript behind the scenes when a user uploads a file.
- Elasticsearch. A log field became an Elasticsearch command (discussion) walks through how a value that was only meant to be logged ended up being executed.
A short checklist for today:
- Find every place GnuPG and Ghostscript run, including container base images and build runners, and update them.
- List every feature that processes files uploaded by users. Where you can, run conversions in a sandbox with no network access.
- Treat log fields, headers, and filenames as untrusted input everywhere they flow. That includes search and analytics systems.
Proofs You Can Check
The math story from the past day has moved from claims to checking.
- The Lean community posted about Lean verification of the Riemann Hypothesis (discussion).
- A separate page hosts the verification files for the Lean proof (discussion).
- Another post, Hurried Announcement: Alessandro Masullo GRH (discussion), adds to the same thread.
Not everyone is convinced. A paper on PhilArchive argues for the invalidity of the Alpoge proposed proof (discussion). For broader reactions, see Scott Aaronson's The stochastic parrots have apparently proved the Riemann Hypothesis (discussion) and The Economist's AI is swallowing maths. Is economics next? (discussion).
For software teams, the lesson is about checking, not number theory. A proof assistant moves trust from the author to the checker. It also leaves one question that people still have to answer: does the formal statement actually say what the headline claims? That is a review problem, and humans still own it.
The same idea applies to ordinary code. Deterministic Simulation Testing Is for Everyone (discussion) argues that ordinary teams can now replay a system under controlled time, network, and randomness. It is no longer a technique reserved for database companies.
Small Models and Sharp Benchmarks
- Thinking Machines released Inkling 2 Mini (discussion). Within a day, someone had a quantized build running Inkling-2-Mini in 5.4 GB of memory (discussion).
- Astra 6.2 and Sol 6.2 flop at Simple Bench (discussion). This is a reminder to run evals on your own tasks before you switch models.
- LLMs can read and write executables byte by byte from a hex dump (discussion).
- ClickHouse introduced Loom, agent memory that keeps up with your code (discussion).
- ClawDOS (discussion) is an AI coding agent that runs on DOS.
Where Your Prompts Go
Two reports in the newest queue matter if you send customer data to model APIs.
- Dragos writes that zero retention AI providers were caught logging prompts. It used canary tokens to detect the logging (discussion).
- A Tell HN post quotes an OpenRouter notice: "Requests you sent to MiniMax models may have been read".
Signal's post on restricting AI agents in mobile operating systems from capturing data (discussion) looks at the same problem from the device side.
Three practical steps follow from these reports:
- Keep a list of which providers receive which data.
- Plant canary strings in test traffic so you can tell if prompts are retained.
- Strip secrets before a prompt leaves your network.
The Job Is Changing
- We promoted developers into architects without informing them (discussion) argues that as agents write more code, engineers spend more of their time on system design, interfaces, and review.
- An Ask HN from an engineer with 15 years of experience asks what to do after two years unemployed. The replies give a candid view of the current market.
- A newer Ask HN asks whether you have changed your job due to AI.
- "I can tell you wrote this yourself" (discussion) makes a case for writing in your own voice.
Builder Reads
| Read | Good for |
|---|---|
| PostgreSQL Extensions in Go (discussion) | Teams that want Postgres extensions without writing C |
| Zero C#: Power of C# without any GC (discussion) | .NET developers chasing predictable latency |
| Probabilistic data structures in Go, part 2 (discussion) | Counting and deduplicating at scale |
| The tech architecture of an art website (discussion) | Small site owners comparing stacks |
| Brazil's tax authority open-sources its tax calculator under MIT (discussion) | Public sector and fintech developers |
| Slopstrap: make your site look hastily vibecoded (discussion) | Anyone who has noticed that AI-built sites look alike |
| DreamCoder (discussion) | Emacs users evaluating coding harnesses |
| Winedroid WASM demo (discussion) | Running Android apps in the browser |
Quick Hits
- My monitor was cutting my download speed by 75%: a Studio Display Ethernet debugging story. (HN)
- Dumping an AES key from a MyQ garage opener to stop reboots when it is offline. (HN)
- MPEG Gaussian Splat bitstreams, decoded in the browser (HN)
- Detonating a pile of lead azide (HN)
- The Soviet Elektronika BK-0010: a home computer built on a PDP-11 clone. (HN)
- An ASCII cell aquarium for your terminal (HN)
- How I made a custom game engine for Satelital (HN)
- NINLIVE: The Nine Inch Nails live archive (HN)
- Sir Brian May adds new dimension to Hera asteroid (HN)
- Surprisingly adorable insect portraits (HN)
- Defcon 5's California Caravan program and zine (HN)
- Clojure will affect the way you think about programming (2017) (HN)
What This Means for Content Teams
Keep file processing off your own servers. The GnuPG and Ghostscript bugs live in the pipelines that convert uploads. With a headless CMS, uploads go through a managed media API and are served from an image CDN. Content reaches your front end as data through the REST API, so your team has fewer conversion tools to patch when the next RCE lands.
Check before you publish. The Lean story shows that a checker earns trust only when someone confirms it checked the right statement. Content works the same way. Cosmic AI Agents work inside your content model and keep activity logs. Cosmic Workflows add approval gates, so an editor reviews AI-assisted drafts before they go live.
Know where your content and keys go. The prompt logging reports are a reminder to limit what leaves your system. Each Cosmic Bucket has separate read and write keys, and the authentication docs explain why the write key should never ship in a browser bundle. For coding assistants, the hosted MCP Server exposes a fixed set of content tools rather than open access.
Design the interface first. If developers are becoming architects, the content model becomes one of the interfaces they own. A clear model lets people, agents, and visiting AI tools all read the same structured content, and a front door for AI agents answers from what you have published.
To get started, read the documentation or create a free account.






