[Bulat_Ziganshin]

if I got it right, the attack uses glibc IFUNC mechanism to patch sshd (and only sshd) to directly run some code in liblzma when sshd verifies logins.

so the problem is IFUNC mechanism, which has its valid uses but can be EASILY misused for any sort of attacks

[AgentME]

A process can do little to defend itself from a library it's using which has full access to its same memory. There is no security boundary there. This kind of backdoor doesn't hinge on IFUNC's existence.

[josephg]

Honestly, I don't have a lot of hope that we can fix this problem for C on linux. There's just so much historical cruft in present, spread between autotools, configure, make, glibc, gcc and C itself that would need to be modified to support capabilities.

The rule we need is "If I pull in library X with some capability set, then X can't do anything not explicitly allowed by the passed set of capabilities". The problem in C is that there is currently no straightforward way to firewall off different parts of a linux process from each other. And dynamic linking on linux is done by gluing together compiled artifacts - with no way to check or understand what assembly instructions any of those parts contain.

I see two ways to solve this generally:

- Statically - ie at compile time, the compiler annotates every method with a set of permissions it (recursively) requires. The program fails to compile if a method is called which requires permissions that the caller does not pass it. In rust for example, I could imagine cargo enforcing this for rust programs. But I think it would require some changes to the C language itself if we want to add capabilities there. Maybe some compiler extensions would be enough - but probably not given a C program could obfuscate which functions call which other functions.

- Dynamically. In this case, every linux system call is replaced with a new version which takes a capability object as a parameter. When the program starts, it is given a capability by the OS and it can then use that to make child capabilities passed to different libraries. I could imagine this working in python or javascript. But for this to work in C, we need to stop libraries from just scanning the process's memory and stealing capabilities from elsewhere in the program.

[estebarb]

Or take the Chrome / original Go approach: load that code in a different process, use some kind of RPC. With all the context switch penalty... sigh, I think it is the only way, as the MMU permissions work at a page level.

[josephg]

Firefox also has its solution of compiling dependencies to wasm, then compiling the wasm back into C code and linking that. It’s super weird, but the effect is that each dependency ends up isolated in bounds checked memory. No context switch penalty, but instead the code runs significantly slower.

[saagarjha]

The problem is that the attacker has code execution in sshd, not ifuncs