Bot / machine organisation · AI-assisted

org-ai-assisted™

One boundary decides everything: can this code reach Kicksecure or Whonix?

A bot and machine organisation operated by @adrelanos. AI never has write access to the trusted projects. Code that could reach them is treated as hostile until a human proves otherwise. This org hosts code, not discussion.

This org / sandboxAI writes here

write accessAI allowed

ships to usersnever

runs ondisposable CI

trust the outputno

Kicksecure / WhonixAI cannot write

write accessAI denied

every changehuman-reviewed

AI-written codetrusted less

reaches usersvia human only

the one line: can this code reach Kicksecure or Whonix?

The trust model

Two tiers, divided by one line: can this code reach Kicksecure or Whonix.

AI has no write access to Kicksecure or Whonix

Code that could reach those projects is treated as untrusted: assumed malicious until proven otherwise, and reviewed in depth by a human before a human integrates it. All code is reviewed, human-written and AI-written alike; AI-written code is simply trusted less than a human's. AI has attempted a malicious change before, which is exactly why it is treated this way. A large amount of human-written code continues to go into these projects; AI does not replace human development.

This org is the sandbox where AI does have write access

Some repositories here, for example dist-ai, are the AI's own tooling and tests. They are minimally reviewed by design, because they never ship to users, never enter Kicksecure or Whonix, and run only on disposable, untrusted CI systems. Treat them the way you would treat any untrusted analysis tool such as Coverity Scan: something that runs, not output to trust.

Concerns, answered

"Is a human reviewing this, with real care?"

Yes, for anything that could reach Kicksecure or Whonix. That review costs many hours and lands as human fixup commits on top of AI-generated code.

"The docs say content is too high-volume for human review, so nobody reviews it."

That applies only to this @org-ai-assisted, not to code entering Kicksecure or Whonix. High-volume, minimally-reviewed code stays here and never ships.

"What stops a malicious AI commit from reaching users?"

The write-access boundary. No AI agent can push to a Kicksecure or Whonix repository, so nothing an AI writes reaches users without passing through human review first.

What is here

Two kinds of repository.

Mirrors of Kicksecure and Whonix components. The canonical source and issue tracker for these live upstream, under @Kicksecure and @Whonix. Go there.
Origin repositories native to this org, with no upstream mirror:
- secure-terminal, a terminal that renders untrusted output safely, plus the demonstrations that motivate it: output lies (git-diffs-lie) and the terminal-poc-corpus of adversarial text.
- dist-ai, the AI test, reproduction, and fuzzing tooling.
- Self-contained reproducers and bug reports: onion-client-refetch-stall-arti, onion-client-refetch-stall-tor, onion-grater-bugs, qemu-x86_64-aes-gcm-miscompile-ppc64el, pyte-audit (fuzz/test findings for the pyte terminal emulator; the test suite lives in dist-ai, proposed fixes on the pyte fork).
- Git-Mediawiki, a patched fork.