[benregenspan]

Is the "Jia Tan" XZ Utils compromise not a good example? That relied on code snuck into a release that was not in source.

(It was caught before being promoted into a stable Debian release, yes, but this sort of relied on a happy accident, too close for comfort)

[chuckadams]

The xz hack was still reproducible, because it was included in the distribution archive which did not match the upstream source -- even then, it was so obfuscated it likely would have gone unnoticed, but nevertheless it only lived in the uploaded tarball and not in the repo. Reproducibility is a good thing, but the next step is build provenance.

Still, lots of good non-security benefits to reproducible builds too.

[jcranmer]

The xz utils compromise is a very good example... of why reproducible packages doesn't actually solve anything security-wise!

The backdoor relied first on a difference between building a package in a packaging environment versus building the package on your own. And also, it relied on the very common practice of checking in unreviewable artifacts into the source tree (e.g., the configure script, malicious binary test artifacts).

Reproducible builds guarantee that two people can follow the same instructions and get the same, bit-identical outcome. It does nothing to guarantee that those instructions have not been compromised, and all of the great packaging security failures of my lifetime that I can think of have relied on those instructions being compromised (e.g., xz utils, Debian OpenSSL keygen issues).

[uecker]

An attack would be far easier without reproducible packages. One could upload a compromised binary to debian by becoming a debian developer, blackmail a debian developer to so, or compromise the computer of a debian developer or the distribution.

At the time of xz attack, the package was already reproducible.

[jcranmer]

I'll give an analogy to email and spam. A lot of effort has been spent making sure that if an email is from [e-mail], it actually came from [e-mail], giving us things like SPF, DKIM, and DMARC. And it turns out that the most eager adopters of the newest technology are... the spammers themselves! Because they don't need to lie about their email address; they can have that be completely honest, and instead resort to other tricks to mislead users as to who they are (e.g., the display name, which most email clients blindly trust and happily display).

Similarly for package managers, the biggest issues are typo-squatting or maintainer credentials compromise. And in neither case does the attacker have any incentive to take advantage of it in a way that breaks reproducibility--they can be completely honest about what they're doing. Now even if I were an attacker who had compromised a maintainer's machine, I'd still probably go for compromising the source rather than compromising the final artifact-generation process... simply because compromising the source makes the exploit live longer.

As xz shows, once you have a compromised maintainer, there's basically nothing you can do to fix it except by having someone else discover the compromise and locking out the maintainer.

[uecker]

Typo-squatting attacks are more of an issue for non-curated software collections, not so much for Debian. If you use npm or cargo or similar, then you have indeed far bigger worries. Compromising the source has the disadvantage for the attacker that it much easier to detect. Again, if you always install the newest things from an non-curated collection that may not matter much, but for something such as Debian, this increases the probability of detection a lot. One can argue that xz shows that it is possible to hide things in the source, but it also shows how much effort was needed to do this. (and the xz package was reproducible, so compromising the debian system and uploading a binary would have a high risk of detection. That this was not done can therefor not serve as evidence that binary attacks are not an issue. )

[goodpoint]

No, reproducible builds make such backdoors more difficult to sneak in together with other checks.