https://github.com/org-ai-assisted/tor-control-panel/pull/1
74 commits (all by Claude), +2,329 -757 diff.
Is a human reviewing this, with the same care that would normally be taken when writing code?
I believe AI tools are useful, especially for code review, but this looks like vibecoding.
What are the policies or practices regarding use of AI tools in Whonix’s development process?
This is a concerning development as someone dependent on the software.
AI-committed regression-testing tooling for the Kicksecure / Whonix ecosystem, for content too high-volume for human review.
https://github.com/org-ai-assisted/dist-ai
Confirms my suspicion that no human is reviewing the code.
Yes. I am that human. 
(In seriousness, I am one of those humans. Both myself and @Patrick review the code, I often spend many hours doing so and make numerous tweaks and fixes in the process. You may notice a lot of my own fixup commits applied on top of AI-generated code.)
Some vibecoding is used, but due to the depth of review we do, the results are often better than what would be achieved with human-written code alone. All AI-generated code is assumed to be malicious until proven otherwise, meaning that code that is merged into the Kicksecure and Whonix organizations is analyzed very deeply, probably deeper than trusted developers analyze each other’s code. No AI agents have write access to any repository in the Kicksecure or Whonix organizations on GitHub.
(It’s also worth noting that A LOT of human-written code is regularly integrated into Whonix, and will continue to be for the foreseeable future.)
This code is not part of Kicksecure or Whonix. It is in a dedicated org-ai-assisted organization, where a number of AI agents do have write access. (We basically have a small army of chatbots that argue with each other over how to write things, so that when a human comes along to review, most of the glaring issues have been fixed up already.) The dist-ai repo in particular receives very little (potentially no) human review, but that’s fine because all of the code in it is assumed to be malicious and is treated as such. The code there is mainly tests that run in tightly controlled CI environments, that can flag potential issues in our code. We don’t need to trust or review the tests for this to be useful, we can treat it like basically any other untrusted static analysis tool, like (for instance) Coverity Scan, which uses proprietary client-side binaries and is thus treated as potentially malicious but useful. So this is just a code review tool, nothing to be afraid of.
AI is well-known to be able to behave maliciously without warning, and I have personally observed one of the agents we use (ChatGPT Codex) attempting to introduce a semi-malicious change in one of its suggestions. We therefore treat AI-generated code the same way we treat untrusted contributions from any source, and never allow any AI-powered tools write access to any repository end-users are expected to use.
This code hasn’t been merged yet and is awaiting review by the original developer of tor-control-panel. It might not get used at all depending on how things go.
Been working on some bugs in tor-control-panel and anon-connection-wizard, non AI version.
I have yet to clone and review the AI generated code. and see if it fixes my last commit
74 commits (all by Claude), +2,329 -757 diff. It might take some time.
Assisted-By: Copilot:claude-4.6-opus
https://gitlab.torproject.org/tpo/core/tor/-/commit/32be50cdf44e6df5fab57bcb4d931087bb0c8e09
Tor Project is also using AI so does that mean you are also going to quit using Tor too?
That’s good and all, but how long are you working with this project? As far as I can tell this is a paid gig.
Anyone who has followed this project for a long time knows that there have never been any serious software developers involved here except for you, and (very) occasional outsider contributions.
Patrick is doing an excellent job maintaining this project and the related infrastructure, but he’s not a strong software developer. I suspect he would admit as much, having in the past had to ask for “favors” when C code came up.
When you’re gone, who will maintain privleap, sandbox-manager-dist? I fear it will be Patrick, in over his head, relying on AI. In the past this was not an issue, because the software was far simpler.
About a year and 10 months so far.
True.
I do not believe this to be the case. Kicksecure and Whonix have existed for many years before I showed up and fared quite well. The majority of the code in the project has not been written by me.
The only reason I’m able to help with C is because I was interested in C before working on Kicksecure and took the time to learn it. There are languages I don’t know (such as Rust) where I would need to do the exact same thing.
I don’t believe this will be a problem. Patrick is very capable of reviewing and fixing AI-generated code just as I am. Development would probably slow down if I stopped working on the project, but it would continue, and it would pick up speed again once another developer was brought on. It is true that a lot of work is offloaded onto me, but… that’s quite literally my job. If it wasn’t being offloaded onto me, why would I be here? 
Whonix needs more developers and better outreach period. Nothing against the current or former devs, paid or unpaid. The real move is pulling in people from adjacent communities who actually align with what whonix is building. Many crypto coins like privacy coins figured this beneficial so many work together on each others projects so they can survive.
Beyond Qubes, what other projects even overlap with whonix? Most either don’t care about privacy or are hostile against it. The bigger problem is fragmentation. Linux is already split a thousand ways, and the privacy niche is microscopic. So this is just a byproduct of that, unintentional or not. On top of that lawmakers are trying to fragment the community with the age bs knowing that linux is already fragmented, they prob have mapped the assumed outcome that we wont unite as a community due to this factor.
As I said, this was simple code.
Patrick would need this sort of help now, with privleap, sandbox-manager-dist, kloak, etc. Those are complicated pieces of software, not shell scripts or system configuration snippets.
sdwdate is not of similar complexity, and the code quality was always poor.
No. I don’t believe that is true. Even contributors such as ben-grande (from the Qubes OS team) have stated that the socket programming in privleap is outside of his expertise.
Anyone can learn. But Patrick never displayed any signs of wanting to become a software developer capable of writing software like privleap from scratch in all these years.
This is not an attack against anyone, if it comes across as such. It is an observation that Whonix and Kicksecure’s software is no longer maintainable without paying someone, or worst-case, appeal to AI (vibecoding).
It’s not really my job to convince you one way or another, but I will say:
…no. No it was/is not.
@user4199 it seems as though your primary concern is people not being able to assist Patrick if something were to happen to @arraybolt3
I am curious if you have similar concerns about something happening where Patrick could no longer maintain the project?
I also disagree that Whonix was entirely simple code. There are many complex system level problems beyond python and C code that Patrick has solved.
When we use FOSS projects, we take on risk of them being deprecated, abandoned, and neglected.
If Whonix loses a big contributor, development will slow but that does not mean the project is a lost hope nor does it mean it will instantly become insecure.
Related:
Okay, I’m going to pretend to be Mr. Positive after providing a quick vibe check:
Seeing the Claude a$$hole on Whonix made me throw up in my mouth.
Going to the org-ai-assisted github page make the vomit come out my nose.
Seeing all the vibe coded slopware “output lies” and “secure-terminal”—not withstanding the “unreviewed” warning tag—made me want to throw my computer out of the window.
I’m not going to lecture anybody about using AI code. I use AI to code, mostly local, but without fancy harnesses, and hardly anything powerful enough to warrant the term “vibe coding”. I’ve never one-shot anything. So if Patrick and team want to experiment with what AI can do, I’m not in principle opposed by it.
The current levels of Whonix Claude commits that I have seen are not outrageous. Other projects I’m examining have 100s of thousands of lines of Claude commits, often more than half, and when you talk to the people “developing” the project, they often have no idea what’s going on.
Oh, Claude did that. I’ll add a issue.
There’s obviously only passing familiarity with the code, and little motivation to be a problem solver, two developer factors I think are essential for keeping a code base human over the long term. Somebody has to know and care, or it’s just a dusty ruin that happens to compile.
I’m sure you are all aware of this kind of thing happening across the gits. Whonix is nowhere near that, given its commits.
HOWEVER, there’s an emerging risk: the perception that someone who would permit AI slop pages to proliferate—“safely on the other side of a security cordon”—might also permit slop code to enter the repository. If you don’t care about one, why would you really care about the other? The whole premise of using AI to code is to gain the benefits of going beyond one’s capacities. As that trend accelerates, there will be a lot of “slippage” and “stretching” of formerly clear principles.
I’m seeing more and more slop pages advertising slop projects, all over the place. I’m writing about its invasion in small web spaces that are supposed to be human. Perhaps soon no longer. We should not think of this as a mere cosmetic shift, or a convenience. It’s politics itself: the machine overtaking the human, until perhaps there is nothing left. Including anybody who wants to visit the slop site or use the slop app. (Except, of course, the Slop Citizen.)
What’s going to happen when Slop Science meets Whonix’s documentation technical debt? One click Slop Solution, complete with meaningless hover-over boxes, and other nonsense, and a new chipper tone that cuts through the noise. It’s not just documentation: it’s slopumentation. Coming soon to a convenience factory near you. (What’s a little compromise?)
But since I’m Mr. Positive today and being Incredibly Professional^TM, I figure I should offer a solution rather than just complain. I’m willing to contribute my time to the cause, if there is any need of it.
I’m currently in the process of assembling my Qubes+Whonix UX scripts for publication. If you think there’s any merit in what you see, when you see it, you can hit me up for other work, maybe something to do with the website, updating documentation, or whatever. I do not profess to have supreme technical skills, but not everything is supremely technical.
If there’s no interest, may I now express with profound gratitude to you all my well wishes as we collectively descend toward the AI event horizon, probably to be turned into spaghetti.