Debian apt security update (man-in-the-middle)
lists.debian.orgI found this! I posted a semi-click-bait-ey tweet about it here [0], I think it’s important for people know how easy these bugs are to find now. It’s becoming borderline irresponsible not to audit your internet-facing code with modern AI tools.
Looks like you also found a similarly critical apt bug in 2019: https://lists.debian.org/debian-security-announce/2019/msg00...
Do you think there might be even more still lurking out there?
And what is your take on HTTPS for apt? I assume the MitM scenario for this and the previous bug could have been avoided by the use of HTTPS by default?
HTTPS would have mitigated this particular bug but it feels like rearranging deck chairs on the titanic to fight over that right now. We need to audit everything with powerful AI like, yesterday.
I'm sure Debian was invited to take part in Project Glasswing back at the beginning of this year.
I wonder if all the people who laughed off Anthropic's sensationalistic claims about Mythos as marketing fluff have recanted, or if they're too busy patching, or maybe they're stochastically parroting Ed Zitron's latest blather.
I do. I think that security-critical but overly complex software like this should be rewritten and certain aspects of it formally verified, since such things are now practical.
Nice work! A similar vulnerability in 2019 also found by you included the recommendation to update with:
Since this also involves redirects, would that protect users while updating apt for this one also?apt -o Acquire::http::AllowRedirect=false update apt -o Acquire::http::AllowRedirect=false upgradeThat sounds right to me, but FYI the APT maintainers themselves are recommending Acquire::Retries=0 as the mitigation since some mirrors might actually have redirects.
I saw the ubuntu announcement also defers to the `apt upgrade -o Acquire::Retries=0` recommendation
This is also a reason why I'm trusting less and less the companies that strongly push against AI. Especially for SaaS type of businesses, AI red teaming is now a critical part of development process.
Speaking of which… if any of y’all work on security at OpenAI I have some questions for you about Daybreak! Please email me
Every security engineer with access to Daybreak (Red or Blue) has a direct communication channel to OpenAI through an assigned Account Executive, or at least your employer does. If you are independent and authorized to use the model, your best option is to email/chat via the support page. The two Account Executives assigned to us are good people but not very technical, so I hope you have better luck with yours.
For those of you not familiar, apt-cacher-ultra is fairly new (~9 months old), and it can reduce your exposure to this. It uses a completely different mechanism for verifying signatures, and does validate signatures, so if your apt clients grab the packages from the cache, it shouldn't be open to this bug.
The next version is going to include TLS and ACME support for the client->cache connection.
dupe (2h ago): https://news.ycombinator.com/item?id=50364814
I feel like they should release manual update instructions since the update mechanism is vulnerable. Although I see my systems already updated via unattended upgrades.
edit: On lobste.rs I see a recommendation for `apt download`: https://lobste.rs/s/h7bvma/man_middle_arbitrary_code_executi...
CVE says you need a non zero retries, so you can use Acquire::Retries=0 to avoid the issue completely (unless your packages were modified already of course).
But how do I update apt if apt is untrustable until fixed ? Good ol' .deb scp'd in dpkg'd.
It always bothered me that apt didn’t use encryption, I know the issue here doesn’t boils down to that, but damn this sucks.
Sure I can update, but how to check I want compromised? Will apt check the state of past/installed updates?
How would an attacker inject a redirect into an apt package download?
MITM on the network, I guess like DNS takeover or the like.
So you pass through the upstream headers but then simulate a failure/resume and supply different contents for the resumed download.
So apt doesn't verify signatures on what is actually downloaded?
> Improper file handling when resuming an HTTP download allows a man-in-the-middle attacker to replace a package after verification
from the OP
APT can use HTTPS, but often doesn't, making a MITM attack easier (also the idea for a long time has been signatures are enough, and/or TLS certs used to be a lot harder to get, but at this point you should really be using both).
Still doesn't make sense to me. So package A is downloaded, signature verified, then package B is downloaded, but somehow the package A file is replaced? Is it something like that?
The bug is not trivial, I asked my AI friend to analyze.
To summarize, APT may have several download sessions in parallel, which may download the same "resources". To save bandwidth, disk storage, these same resources are downloaded only one, but given different names on the disk through hard links.
Because of HTTP retry, APT may be lured into downloading again a resource, but that was associated with a different package name, package that was already checked (hash verified). This is because of some differences between the package name / filename / URI, etc.
Anyway, the point is that hash are verified when a package is completed, but because of possible parallel downloads and hard links usage, a verified package can be rewritten.
What I find crazy though is that why packages are not first completely all downloaded, then every time a package is used, the hash be verified before being used. This would reduce even further the surface of attacks, and I don't care about any performance impact.
> What I find crazy though is that why packages are not first completely all downloaded, then every time a package is used, the hash be verified before being used. This would reduce even further the surface of attacks, and I don't care about any performance impact.
Yes I'm amazed to learn this is not the case. Without that, any local exploit that could replace a file could potentially replace a downloaded, "already verified" package.
The somehow is package B has the same filename as package A.
Https doesn't help with this if the MitM is the apt mirror you are using.
One reason http is still used is because you need an independent signature anyway since it is common to use mirrors that don't have the same level of trust as the debian maintainers
We are living in times where I read first in HN what apticron is going to send me by mail.
Discussion here: https://news.ycombinator.com/item?id=50364814