I'm investigating an AI agent platform as part of an bug bounty and would like some input from people experienced with Firecracker, Linux isolation and cloud security.
Here are the observations so far:
The agent terminal initially runs as an unprivileged sandbox user (UID 1002). sudo requests a password.
However, su - root successfully provides a root shell.
As root, capsh --print shows all capabilities in the bounding, permitted and effective sets, including CAP_SYS_ADMIN and CAP_SYS_PTRACE .
The environment runs Linux 6.1.158 and systemd-detect-virt identifies it as KVM.
I can reach 169.254.169.254 and obtain HTTP 200 responses from its metadata endpoints. The server identifies itself as Firecracker API.
I noticed an address metadata entry pointing to 192.0.2.1/logs .…