This software is currently considered unfinished and experimental.
It will remain so until anyone starts using it for anything but experiments.
Until people actually start using this software in production and sharing traces, there is no reason to be coy about any security issues you might find. Just open a regular issue or merge request in our github project.
There are two usage patterns for this software: Capture and replay (replay including here post-processing or inspect).
This software has no privileges that the captured app does not already have during capture, so there is nothing to be gained from securing it.
Replay (parsing the trace file in any way with this software) can, however, introduce an untrusted input into a trusted environment if the trace file comes from outside the trusted environment (eg a customer or an outside bug report), and this carries with it the risk of a security breach.
API tracing exposes a very large attack surface to tools that can be exploited by a clever attacker by modifying a trace file by hand and sending it to an unsuspecting victim. Instead of attempting to harden this entire surface entrypoint by entrypoint, lavatube is experimenting with a sandbox to reduce what can be gained by such an attack.
This sandbox is composed of four levels:
- Zero - sandbox disabled.
- One - ability of getting root privileges removed.
- Two - network and file execution privileges removed. Individual tools may retain outgoing TCP access to explicitly allowlisted destination ports.
- Three - file system write privileges removed except to current working directory and directories underneath it.
The first level is entered on program startup, the second after parsing command line arguments, and the third just before parsing the trace file. This allows us to gradually drop privileges as as they are no longer needed.
However, Vulkan layers are also affected by these capability changes, so validation layers, screenshot layers, and performance tools might all become dysfunctional.
lava-agent, lava-cli, and lava-tui need to create new connections while
running. At sandbox levels two and three, they retain outgoing TCP access only
to their configured destination ports. This is the replay-service port for
lava-cli, the configured model and optional replay-service ports for
lava-tui, and the model and replay-service ports for lava-agent. Landlock
restricts ports rather than remote addresses, so this does not limit
connections to a specific host on an allowed port.