I still believe eBPF, if not managed correctly, is a big security issue. Containers isolate processes, not the kernel. eBPF runs in the kernel.
Netfilter was fixed for this: a container's firewall rules only apply to its own network namespace. Scoped by design.
eBPF tracing hooks weren't. A kprobe loaded from one container sees every process on the host. CAP_BPF in a container is really a host-wide grant.
Anyway, my 2 cents.