Advanced container security

Correct — there is no container object — just a host process in a costume.

Correct — it can make container root a harmless host user, breaking root-equals-root.

Correct — they must be paired with capabilities, seccomp, and ideally a user namespace.

Correct — without limits one container can take down the host.

Correct — set --memory-swap equal to --memory (or disable swap on container hosts).

Correct — so a process gets only the specific powers it needs.

Correct — mount alone has been chained into container breakouts.

Correct — usually none, sometimes just NET_BIND_SERVICE.

Correct — why a breakout as container root is a host takeover.

Correct — the single strongest mitigation for root-equals-root.

Correct — the runtime image ships only the artifact — far less to attack.

Correct — most post-exploitation tools are simply absent.

Correct — the attacker is limited to what your binary itself can do.

Correct — tools come alongside at runtime; the production image stays inert.

Correct — so an empty base works: smallest image, zero OS surface.

Correct — scratch has no trust store; you add what the binary assumes.

Correct — Go/Rust static → scratch; most else → distroless.

Correct — cleanup only helps inside the same RUN that created the file.

Correct — the SBOM lists components so a scanner can match CVEs without an on-disk DB.

Correct — minimal reduces surface; provenance proves contents and source.

Correct — it turns a breakout from host takeover into an unprivileged nuisance.

Correct — runAsNonRoot checks the UID and cannot resolve a name to 0.

Correct — every writable location becomes declared and auditable.

Correct — kills a whole class of escalation for free.

Correct — root-with-no-caps and non-root-with-SYS_ADMIN each leave a path.

Correct — verify Seccomp: 2 in /proc/self/status in production.

Correct — deny-by-default plus an observed allowlist.

Correct — seccomp filters syscalls; the LSM confines resource access.

Correct — complain/permissive mode exists to tune without turning it off.

Correct — a quick way to catch a container running without seccomp.

Correct — a daemon or breakout compromise is a user compromise, not host root.

Correct — no central root daemon to attack; containers are children of the CLI.

Correct — the container asks the daemon to hand over the host.

Correct — use mutual TLS on 2376, firewall it, or do not expose it.

Correct — socket access can mount / and chroot as root.

Correct — a safer default so containers cannot freely reach each other.

Correct — it scores dozens of config checks and flags drift.

Correct — root inside, harmless user outside.

Correct — its writes land as the mapped unprivileged UID.

Correct — sharing the host user namespace removes the mapping.

Correct — which means disciplined config prevents most of them.

Correct — effectively a host root shell with extra steps.

Correct — scope the exact privilege instead of granting everything.

Correct — the container did not escape the kernel; it used the API.

Correct — e.g. docker-socket-proxy; or use rootless DinD / Kaniko so no socket is needed.

Correct — each leaks credentials or kernel interfaces or a persistence path.

Correct — exactly what cap-drop ALL removes.

Correct — the shared kernel is the ceiling; these raise the bar and contain the unknown.

Correct — a kernel-level escape lands in the sandbox, not on the host.

Correct — hardening turns the workload into a sharp tripwire.

Correct — segment with user-defined/internal networks and default-deny egress.

Correct — no route anywhere — nothing to pivot to or exfiltrate through.

Correct — a compromised container or SSRF can assume the node’s role and pivot into the account.

Correct — so your allow/deny survives Docker rewriting its chains.

Correct — anyone who can look at the container can read it.

Correct — in memory, read-only, absent from inspect and the environment.

Correct — a leaked credential expires on its own and rotation happens in one place.

Correct — host /proc exposes secrets in environ and open FDs.

Correct — each reopens a wall the rest of your hardening built — like --privileged.

Correct — contain first, then capture with diff/export/commit, then investigate.

Correct — a well-behaved container changes little, so attacker artifacts stand out.

Correct — assume all reachable secrets are compromised; never restart the popped container.