Compiled software is no longer a safe hiding place for source code.
Using AI agents and only the shipped binaries and public documentation,
we reconstructed maintainable source code for Slack (client and server),
Apple's Finder, and the TinyTurla-NG backdoor, then added a feature Slack
does not have. We look at how far each reconstruction got and what
vendors can do to protect their software.
Laws around AI right now are such that anything AI produces is not copyrightable.
Companies that push it out can be sued but the plaintiff won’t be able to prevail without a way to pierce the non-copyright framework.
And since the company’s that would be suing benefit from the framework, I suspect it will be a settlement to purchase the code and then remove it from repositories.
That’s not the issue. The problem here is that the article claims “everything is open-source”. But it isn’t, because AI code is tainted: it’s impossible to prove it wasn’t trained on copyrighted data (EXISTING, PRE-AI copyrighted data) and it didn’t use it, copy it or adapt it to create the vibe-coded slop. Tainted code can’t be released as open-source.
Non-copyrighted code, which is all code produced by AI by law, is defacto open to being Open Sourced, since there’s no other option.
It doesn’t matter what it was trained on. The law, right now, as has been proven repeatedly in court, is that all AI-produced products are not able to be copyrighted, but are legal to be generated, even when they are clearly trained on copyrighted data.
It would require a new law to stop it, and if that happens then all AI goes away and all previously released ai-produced code would still be legal since you cannot punish people retroactively.