Shai-Hulud Themed Malware Found in the PyTorch Lightning AI Training Library

Any suggestions for a vm/container setup that works on a Linux host, provides the safety net you describe, and is still capable enough to try out all these things that people are talking about?

Historically, extra-security-scanned artefact handling has been a paid enterprise option. Whereas the less secure option is the much-less-hassle default.

IDK how good a business model this is, I suspect not very.

More like hiding their heads in the sand in circumstances that are outside of their ability to fix. None of the tooling or practices out there push you in the direction of not being at risk, or even provide you with easy ways to stay completely safe: no external packages needed to develop software with everything you NEED being provided out of the box, or a flow where pulling in a new package makes you review all of its source code line by line and compile everything instead of any binary tooling blobs, or built in vulnerability and configuration scanning so you don't get pwned by Trivy or don't leave an open S3 bucket somewhere, which also means that obviously you'd need thorough observability and alerting for any of the cloud stuff you do.

And even when they exist, your org projects might be painfully out of date, too much to use those approaches, or the org culture might not be there, or any number of other issues I can't even imagine. On one hand, people are running out of date software and those have CVEs, on the other using dependencies that are too new also puts you at risks of compromised packages - it's like we're being squeezed by rocks on both sides in a landslide or something. Even at the OS level, the fact that everyone is not running something like Qubes OS or regular VMs for development is absolutely insane. The fact that all software isn't sandboxed and that desktop OSes don't prompt for permissions like mobile apps do is absolutely insane. That we don't have firewalls like Glasswire as standard that prompt you for external connections, or don't allow easily blocking what you don't trust is insane.

Despite lots of people trying their best, on some level, everything both up and down the stack is absolutely fucked for a variety of complex reasons. You'd have to largely tear it all down and rebuild everything starting with your OS kernel in a memory safe language and formal proofs and thorough testing for everything (if it took SQLite as long as it did to get a decent test suite, it might as well take on the order of decades to do it for a production OS kernel and drivers), then do the same for all userland software and DBs and tooling and dependency management and secrets management (not just random files, special hardware most likely) and so on. It's not happening, so we just build towers of cards.

For something more practical: https://nesbitt.io/2026/03/04/package-managers-need-to-cool-...