Now Debian insists we add the GPG keys manually. Like cavemen.)
Erm. Would you rather have debian auto-trust a bunch of third party people? It’s up to the user to decide whose keys they want on their system and whose packages they would accept if signed by what key.
Not “auto trust”, of course, but rather make adding keys is a bit smoother. As in “OK, there’s this key on the web site with this weird short hex cookie. Enter this simple command to add the key. Make sure signature it spits out is the same on the web page. If it matches, hit Yes.”
And maybe this could be baked somehow to the whole APT source adding process. “To add the source to APT, use apt-source-addinate https://deb.example.com/thingamabob.apt. Make sure the key displayed is 0x123456789ABC by Thingamabob Team with received key signature 0xCBA9876654321.”
sudo apt-key adv --keyserver keyserver.ubuntu.com --recv-keys 00000000
where 00000000 is replaced with the fingerprint of the key you want to fetch?
I do agree - the apt-key command is kinda dangerous because it imports keys that will be generally trusted, IIRC. So a similar command to fetch a key by fingerprint for it to be available to choose as signing keys for repositories that we configure for a single application (suite) would be nice.
I always disliked that signing keys are available for download from the same websites that have the repository. What’s the point in that? If someone can inject malicious code in the repository, they sure as hell can generate a matching signing key & sign the code with that.
Hence I always verify signing keys / fingerprints against somewhat trustworthy third parties.
What we really need though is a crowdsourced, reputation-based code review system. Where open source code is stored in git-like versioning history, and has clear documentations for each function what it should and should not do. And a reviewer can pick as little as an individual function and review the code to confirm (or refute) that the function
does exactly what the interface documentation claims it does
does nothing else
performs input validation (range checks etc)
is well-written (in terms of performance)
Then, your reputation score would increase according to other users concurring with your assessment (or decrease if people disagree), and your reputation can be used as a weighting factor in contributing to the “review thoroughness” of a code module that you reviewed.
E.g.: a user with a reputation of 0.5 confirms that a module does exactly what it claims to do:
Module gets review count +1, module gets new total score of +0.5, new total weight of ( combined previous weights + 0.5 ) and the average review score is “reviews total score” / “total weight”.
Something like that. And if you have a reputation of “0.9”, the review count goes +1, total score +0.9, total weight +0.9 (so the average score stays between 0 and 1).
Independent of the user reputation, the user’s review conclusion is stored as “1” (= performs as claimed) or “0” (= does not perform as claimed) for this module.
Reputation of reviewers could be calculated as the sum of all their individual review scores (at the time the reputation is needed), where the score they get is 1 minus the absolute difference between the average review score of a reviewed module and their own review conclusion.
E.g.
User A concludes: module does what it claims to do: User A assessment is 1 (score for the module)
User B concludes: module does NOT what it claims to do: User B assessment is 0 (score)
Module score is 0.8 (most reviewers agreed that it does what it claims to do)
User A reputation gained from their review of this module is 1 - abs( 1 - 0.8 ) = 0.8
User B reputation gained from their review of this module is 1 - abs( 0 - 0.8 ) = 0.2
If both users have previously gained a reputation of 1.0 from 10 reviews (where everyone agreed on the same assessment, thus full scores):
User A new reputation: ( 1 * 10 + 0.8 ) / 11 = 0.982
User B new reputation: ( 1 * 10 + 0.2 ) / 11 = 0.927
The basic idea being that all modules in the decentralized review database would have a review count which everyone could filter by, and find the least-reviewed modules (presumably weakest links) to focus their attention on.
If technically feasible, a decentralized database should prevent any given entity (secret services, botfarms) to falsify the overall review picture too much. I am not sure this can be accomplished - especially with the sophistication of the climate-destroying large language model technology. :/
Erm. Would you rather have debian auto-trust a bunch of third party people? It’s up to the user to decide whose keys they want on their system and whose packages they would accept if signed by what key.
Not “auto trust”, of course, but rather make adding keys is a bit smoother. As in “OK, there’s this key on the web site with this weird short hex cookie. Enter this simple command to add the key. Make sure signature it spits out is the same on the web page. If it matches, hit Yes.”
And maybe this could be baked somehow to the whole APT source adding process. “To add the source to APT, use
apt-source-addinate https://deb.example.com/thingamabob.apt. Make sure the key displayed is0x123456789ABCbyThingamabob Teamwith received key signature0xCBA9876654321.”For the keys - do you mean something like
I do agree - the apt-key command is kinda dangerous because it imports keys that will be generally trusted, IIRC. So a similar command to fetch a key by fingerprint for it to be available to choose as signing keys for repositories that we configure for a single application (suite) would be nice.
I always disliked that signing keys are available for download from the same websites that have the repository. What’s the point in that? If someone can inject malicious code in the repository, they sure as hell can generate a matching signing key & sign the code with that.
Hence I always verify signing keys / fingerprints against somewhat trustworthy third parties.
What we really need though is a crowdsourced, reputation-based code review system. Where open source code is stored in git-like versioning history, and has clear documentations for each function what it should and should not do. And a reviewer can pick as little as an individual function and review the code to confirm (or refute) that the function
Then, your reputation score would increase according to other users concurring with your assessment (or decrease if people disagree), and your reputation can be used as a weighting factor in contributing to the “review thoroughness” of a code module that you reviewed. E.g.: a user with a reputation of 0.5 confirms that a module does exactly what it claims to do: Module gets review count +1, module gets new total score of +0.5, new total weight of ( combined previous weights + 0.5 ) and the average review score is “reviews total score” / “total weight”.
Something like that. And if you have a reputation of “0.9”, the review count goes +1, total score +0.9, total weight +0.9 (so the average score stays between 0 and 1).
Independent of the user reputation, the user’s review conclusion is stored as “1” (= performs as claimed) or “0” (= does not perform as claimed) for this module.
Reputation of reviewers could be calculated as the sum of all their individual review scores (at the time the reputation is needed), where the score they get is 1 minus the absolute difference between the average review score of a reviewed module and their own review conclusion.
E.g. User A concludes: module does what it claims to do: User A assessment is 1 (score for the module) User B concludes: module does NOT what it claims to do: User B assessment is 0 (score)
Module score is 0.8 (most reviewers agreed that it does what it claims to do)
User A reputation gained from their review of this module is 1 - abs( 1 - 0.8 ) = 0.8 User B reputation gained from their review of this module is 1 - abs( 0 - 0.8 ) = 0.2
If both users have previously gained a reputation of 1.0 from 10 reviews (where everyone agreed on the same assessment, thus full scores):
User A new reputation: ( 1 * 10 + 0.8 ) / 11 = 0.982 User B new reputation: ( 1 * 10 + 0.2 ) / 11 = 0.927
The basic idea being that all modules in the decentralized review database would have a review count which everyone could filter by, and find the least-reviewed modules (presumably weakest links) to focus their attention on.
If technically feasible, a decentralized database should prevent any given entity (secret services, botfarms) to falsify the overall review picture too much. I am not sure this can be accomplished - especially with the sophistication of the climate-destroying large language model technology. :/