Repository navigation
Feedback on evaluating runsc vs runc for an untrusted-script research PoC #17247
Unanswered
Maciej-Mitura
asked this question in
Q&A
Replies: 1 comment 1 reply
|
Looks good. I would consider packaging the "restricted host gateway" as a gVisor gofer. See the fsgofer and netgofer as example. This lets you shorten application-sentry-gofer-hostkernel roundtrips, and gives you very precise control over what to do with the network traffic to a degree that may be more difficult to do at a rutnime-agnostic level. Regarding question (3):
|
1 reply
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
Hi,
I'm working on a Bachelor research project about defense-in-depth execution of dynamically loaded / potentially untrusted scripts.
The current PoC executes JavaScript/TypeScript in fresh Docker containers using Deno. The surrounding architecture already includes scoped Deno permissions, Docker resource/network restrictions, a restricted Host Gateway for controlled interaction with an active host application, and execution evidence collection.
gVisor is intended to be an additional isolation layer in that architecture.
The relevant execution path is roughly:
untrusted script → Deno → Docker → runc/runsc → restricted Host Gateway
At the moment the project runs with the standard Docker runtime. Support for selecting runsc is already prepared, but I still need to move the experiment to a Linux Docker host where gVisor is actually available.
My current plan is to run the same workload under both runc and runsc, while keeping the rest of the environment as equal as possible.
I currently want to compare:
-basic JavaScript/TypeScript compatibility
The goal is not to prove that gVisor makes arbitrary code "safe". I want to evaluate whether it can add another meaningful isolation boundary to the existing defense-in-depth architecture, and what compatibility/performance trade-offs that introduces.
I would really appreciate feedback from people familiar with gVisor on a few points:
Repository:
https://github.com/Maciej-Mitura/ResearchProject_Hoatloading_Scripts
The project is still an active research PoC, so criticism of the experiment design is very welcome.
All reactions