diff --git a/README.md b/README.md
index 355545e0..beb1e920 100644
--- a/README.md
+++ b/README.md
@@ -110,6 +110,9 @@ analysis.
 The commands below will dump the JSON results to `/tmp/results`
 and full logs to `/tmp/dockertmp`.
 
+For a detailed, step-by-step walkthrough on macOS (prerequisites, a worked
+example, and troubleshooting), see [docs/local-analysis.md](docs/local-analysis.md).
+
 [gVisor]: https://gvisor.dev/
 
 ### Live package
diff --git a/cmd/analyze/main.go b/cmd/analyze/main.go
index ff5b2fd0..57ac79f8 100644
--- a/cmd/analyze/main.go
+++ b/cmd/analyze/main.go
@@ -131,9 +131,13 @@ func makeSandboxOptions() []sandbox.Option {
 }
 
 func dynamicAnalysis(ctx context.Context, pkg *pkgmanager.Pkg, resultStores *worker.ResultStores) {
-	if !*offline {
-		sandbox.InitNetwork(ctx)
-	}
+	// Dynamic analysis always starts a packet capture on the cni-analysis
+	// bridge (see internal/dynamicanalysis), so the bridge must exist even in
+	// offline mode. When offline, the sandbox container itself still runs with
+	// --network=none and produces no traffic, so the capture simply records
+	// nothing. Previously InitNetwork was skipped when offline, which left the
+	// bridge absent and caused packet capture to fail before any phase ran.
+	sandbox.InitNetwork(ctx)
 
 	sbOpts := append(worker.DynamicSandboxOptions(), makeSandboxOptions()...)
 
diff --git a/docs/local-analysis.md b/docs/local-analysis.md
new file mode 100644
index 00000000..892c4846
--- /dev/null
+++ b/docs/local-analysis.md
@@ -0,0 +1,321 @@
+# Analyzing a local package on macOS (including Apple Silicon)
+
+This is a step-by-step tutorial for scanning a **local-only** package — a package
+archive sitting on your disk that has *not* been published to a registry — using
+the Package Analysis tooling on macOS, **built and run natively** (no x86
+emulation) on Apple Silicon (M1/M2/M3/M4) as well as Intel Macs.
+
+By the end you will have built the analysis images for your Mac's architecture,
+built a sample (deliberately malicious) Python package, run static and dynamic
+analysis on it inside a gVisor sandbox, and read the JSON results describing what
+the package did.
+
+> New to the project? Read the top-level [README](../README.md) first for the
+> big picture, then come back here.
+
+## How local analysis works
+
+The supported entry point is the [`scripts/run_analysis.sh`](../scripts/run_analysis.sh)
+helper. It runs the `gcr.io/ossf-malware-analysis/analysis` Docker image, which
+in turn uses `podman` to launch a nested, [gVisor](https://gvisor.dev/)-sandboxed
+container that *detonates* the package. While the package installs and imports,
+the sandbox captures system calls (file/command activity), network connections,
+and DNS queries.
+
+```
+run_analysis.sh  ->  docker run … analysis  ->  podman + gVisor sandbox  ->  JSON results on your Mac
+```
+
+When you pass `-local <path>`, the script mounts your archive into the container
+and analyzes it directly instead of downloading anything from a registry.
+
+## Why you must build the images on Apple Silicon
+
+The **public** images on `gcr.io/ossf-malware-analysis` are published for
+`linux/amd64` only. On an Apple Silicon Mac, Docker Desktop runs an **arm64**
+Linux VM, so those amd64 images run under qemu emulation. gVisor's `runsc` boots
+its own kernel and intercepts the package's system calls, and syscall
+interception *inside* an already-emulated container does not work — so `runsc`
+fails to boot and you get:
+
+```
+ERROR  Dynamic analysis aborted  {"error": "sandbox failed (error starting container: exit status 125)"}
+```
+
+The fix is to **build the images natively for `linux/arm64`**. Docker Desktop's
+arm64 Linux VM then runs them without emulation, and gVisor — which fully
+supports arm64 — boots normally. This branch contains the small changes needed
+for the images to build on arm64 (an architecture-aware AWS CLI download, an
+arm64 PowerShell install, and a fix so `-offline` dynamic analysis works); see
+[What changed for arm64](#what-changed-for-arm64) below.
+
+On an **Intel Mac** the public amd64 images run natively, so you can either use
+them as-is or follow the same build steps below (everything works either way).
+
+## Prerequisites (macOS)
+
+### 1. Docker Desktop
+
+- Install [Docker Desktop](https://www.docker.com/products/docker-desktop/) and make sure it is **running** before you start.
+- Give it enough resources (Docker Desktop → Settings → Resources): **≥ 6 GB memory** and **plenty of disk** are recommended — the dynamic sandbox image alone is ~3.3 GB, and you are running a container-inside-a-container.
+- File sharing must include the directory holding your package and `/tmp`. On macOS `/tmp` is a symlink to `/private/tmp`, which Docker Desktop shares by default, so the default output locations work out of the box.
+
+### 2. Go is NOT required on the host
+
+The Go binaries are compiled **inside** the Docker image build, so you do not
+need Go installed on your Mac. You only need Docker and the command-line tools
+below.
+
+### 3. GNU command-line tools (required)
+
+`run_analysis.sh` is written for GNU/Linux userland. In particular the `-local`
+code path calls `realpath -m`, and the BSD `realpath` that ships with macOS does
+**not** understand `-m` — so on a stock Mac the script fails with
+`realpath: illegal option -- m`. Install the GNU versions with
+[Homebrew](https://brew.sh/):
+
+```bash
+brew install coreutils findutils util-linux
+```
+
+Then put the GNU tools ahead of the BSD ones on your `PATH` (add to `~/.zshrc` to
+make it permanent):
+
+```bash
+export PATH="/opt/homebrew/opt/coreutils/libexec/gnubin:$PATH"
+export PATH="/opt/homebrew/opt/findutils/libexec/gnubin:$PATH"
+export PATH="/opt/homebrew/opt/util-linux/bin:$PATH"
+```
+
+> On Intel Macs, Homebrew lives under `/usr/local` instead of `/opt/homebrew` —
+> adjust the paths accordingly.
+
+Verify GNU `realpath` is now in front:
+
+```bash
+realpath --version   # should print "realpath (GNU coreutils) …"
+```
+
+`findmnt` (from `util-linux`) is optional — the script falls back gracefully if
+it is missing.
+
+### 4. Clone the repo
+
+```bash
+git clone https://github.com/ossf/package-analysis.git
+cd package-analysis
+```
+
+## Step 1 — Build the images for your Mac's architecture
+
+From the project root:
+
+```bash
+make build/sandbox/dynamic_analysis
+make build/sandbox/static_analysis
+make build/image/analysis
+```
+
+(Or just `make build`, which additionally builds the scheduler image you don't
+need for local analysis.)
+
+`docker build` on your Mac targets your native architecture by default, so on
+Apple Silicon this produces `linux/arm64` images. The dynamic sandbox build is
+the long pole — it installs Node, Python, Ruby, Rust, PHP, PowerShell, etc., and
+takes several minutes the first time.
+
+Confirm the architecture:
+
+```bash
+for img in analysis dynamic-analysis static-analysis; do
+  printf "%-18s " "$img"
+  docker image inspect "gcr.io/ossf-malware-analysis/${img}:latest" --format '{{.Os}}/{{.Architecture}}'
+done
+# analysis           linux/arm64
+# dynamic-analysis   linux/arm64
+# static-analysis    linux/arm64
+```
+
+## Step 2 — Make the local sandbox images available to the nested podman (macOS)
+
+This is the one macOS-specific wrinkle. The `analysis` container runs `podman`
+*inside* itself to launch the sandbox containers, and that nested podman reads
+its image store from `/var/lib/containers`, which `run_analysis.sh` bind-mounts
+from the host. Because your sandbox images are built locally (not pushed to a
+registry), you must run with `-nopull` and pre-load them into that store.
+
+On Linux this is done with `make sync` (`sudo podman load` on the host), but
+macOS has no host podman. Two extra problems appear if you try a plain host
+directory: it is shared into the VM over **virtiofs**, on which podman's overlay
+storage driver cannot create whiteouts (`kernel does not support overlay fs`).
+
+The solution is to back the podman store with a **Docker named volume** (which
+lives on the Docker Desktop VM's own filesystem, where overlay works) and load
+the images into it once. A helper script does exactly that:
+
+```bash
+scripts/load_sandbox_images_macos.sh
+```
+
+This saves the locally-built sandbox images, creates a named volume
+`pa-containers`, and loads + retags them into the nested podman store. It only
+needs to be re-run when you rebuild a sandbox image.
+
+`run_analysis.sh` treats a `CONTAINER_DIR_OVERRIDE` value with **no leading
+slash** as a Docker named volume, so the runs below point at `pa-containers`.
+
+## Step 3 — Get a package to scan
+
+The repo ships a sample Python package that simulates malicious behavior (it
+tries to exfiltrate data over HTTPS and read credential files), which makes it a
+great subject because the analysis will actually find something.
+
+Build it:
+
+```bash
+cd sample_packages
+make build_sample_python_package
+cd ..
+```
+
+This produces a `.tar.gz` in `sample_packages/sample_python_package/output/`:
+
+```bash
+ls sample_packages/sample_python_package/output/*.tar.gz
+# sample_packages/sample_python_package/output/sample_python_package-0.0.1.tar.gz
+```
+
+The package name is `sample_python_package` (from its `pyproject.toml`) — you'll
+pass that as `-package` below.
+
+> **Using your own package?** Point `-local` at any archive and set `-ecosystem`
+> to match: `.whl`/`.tar.gz` for `pypi`, `.tgz` for `npm`, `.gem` for `rubygems`,
+> `.zip` for `packagist`, `.crate`/`.tar.gz` for `crates.io`. The `-package`
+> value should be the package's real name (used during the import phase).
+
+## Step 4 — Run the analysis
+
+Set up the environment once per shell (GNU tools on `PATH`, named-volume store):
+
+```bash
+export PATH="/opt/homebrew/opt/coreutils/libexec/gnubin:$PATH"
+export CONTAINER_DIR_OVERRIDE=pa-containers
+```
+
+Then run. `-nopull` is **required** so the script uses your locally-built images
+instead of fetching the public amd64 ones. `-nointeractive` avoids needing a TTY.
+
+```bash
+scripts/run_analysis.sh -nointeractive -nopull \
+  -ecosystem pypi \
+  -package sample_python_package \
+  -local sample_packages/sample_python_package/output/sample_python_package-0.0.1.tar.gz
+```
+
+This runs **both** static and dynamic analysis. When it finishes you'll see a
+"Finished analysis" summary listing where each result was written.
+
+### Online vs. offline (a real trade-off)
+
+- **Default (online):** the sandbox has network access, isolated by gVisor plus
+  iptables rules that block private IP ranges. This is the path the project's own
+  tests use. The package runs to completion and its real DNS/connection behavior
+  is captured — for the sample you'll see it reach out to `www.httpbin.org`.
+  Untrusted code does run with (contained) network here.
+
+- **`-offline`:** disables the sandbox's network entirely (`--network=none`), so
+  nothing can leave your machine — the safest way to detonate untrusted code.
+  System calls, file activity, and executed commands are still captured. The
+  catch: package managers can no longer download dependencies, so an install that
+  needs the registry will fail at the install phase (e.g. pip: `Temporary failure
+  in name resolution` → `No matching distribution found for setuptools`), and
+  later phases won't run. Use offline for packages that vendor their dependencies
+  or when install-time behavior is all you need.
+
+  ```bash
+  scripts/run_analysis.sh -nointeractive -nopull -offline \
+    -ecosystem pypi -package sample_python_package \
+    -local sample_packages/sample_python_package/output/sample_python_package-0.0.1.tar.gz
+  ```
+
+- **`-fully-offline`:** additionally disables the analysis **container's** network.
+  Requires `-local` **and** `-nopull` (which you're already using).
+
+You can also restrict to a single mode with `-mode static` or `-mode dynamic`.
+
+## Step 5 — Read the results
+
+Results are written to these directories on your Mac (created automatically):
+
+| Directory | Contents |
+|-----------|----------|
+| `/tmp/results` | Dynamic analysis JSON (syscalls, network, commands) + execution log |
+| `/tmp/staticResults` | Static analysis JSON (code signals) |
+| `/tmp/writeResults` | Captured file-write contents |
+| `/tmp/analyzedPackages` | A copy of the analyzed package |
+| `/tmp/straceLogs` | Raw strace logs |
+| `/tmp/dockertmp` | Debug logs (mounted as the container's `/tmp`) |
+
+Inspect the dynamic results (Python's `json.tool` is built in on macOS):
+
+```bash
+cat /tmp/results/results.json | python3 -m json.tool | less
+```
+
+For the sample, an online dynamic run captures its attempted exfiltration — note
+the socket and DNS query to `www.httpbin.org`:
+
+```bash
+python3 - <<'PY'
+import json
+d = json.load(open('/tmp/results/results.json'))
+for phase, v in d['Analysis'].items():
+    print(phase,
+          '| sockets:', len(v.get('Sockets') or []),
+          '| dns:', len(v.get('DNS') or []),
+          '| files:', len(v.get('Files') or []),
+          '| commands:', len(v.get('Commands') or []))
+PY
+# execute | sockets: 9 | dns: 1 | files: 238 | commands: 2   (e.g. 35.x.x.x:443 www.httpbin.org)
+```
+
+For an explanation of every field, see [docs/data_schema.md](data_schema.md).
+
+## What changed for arm64
+
+This branch includes three small changes so the toolchain builds and runs
+natively on arm64. They are no-ops on amd64:
+
+- **`sandboxes/dynamicanalysis/Dockerfile` — AWS CLI download.** The hard-coded
+  `awscli-exe-linux-x86_64.zip` URL is now `…-$(uname -m).zip`, so the correct
+  installer is fetched on both `x86_64` and `aarch64`.
+- **`sandboxes/dynamicanalysis/Dockerfile` — PowerShell.** Microsoft's apt repo
+  only publishes an amd64 PowerShell `.deb`, which breaks the dependency solver
+  on arm64. PowerShell is now installed via apt on amd64 (unchanged) and from the
+  official `linux-arm64` tarball on arm64.
+- **`cmd/analyze/main.go` — offline dynamic analysis.** Dynamic analysis always
+  starts a packet capture on the `cni-analysis` bridge, but `InitNetwork` (which
+  creates that bridge) was skipped when `-offline` was set, so offline dynamic
+  analysis failed before any phase ran (`failed to start packet capture …
+  no such network interface`). The bridge is now created for dynamic analysis
+  even when offline; the sandbox itself still runs with `--network=none`.
+
+## Troubleshooting
+
+| Symptom | Cause | Fix |
+|---------|-------|-----|
+| `realpath: illegal option -- m`, or `-local` seems ignored | macOS BSD `realpath` lacks `-m` | Install GNU coreutils and fix `PATH` (see [Prerequisites](#3-gnu-command-line-tools-required)) |
+| `sandbox failed (error starting container: exit status 125)` | gVisor `runsc` can't boot under amd64 emulation | Build the images natively (`make build`) and run with `-nopull`; don't use the public amd64 images on Apple Silicon |
+| `kernel does not support overlay fs: unable to create kernel-style whiteout` | Nested podman store on a virtiofs host-dir bind mount | Use the named-volume approach (`scripts/load_sandbox_images_macos.sh` + `CONTAINER_DIR_OVERRIDE=pa-containers`) |
+| `failed to start packet capture (… cni-analysis: no such network interface)` | Old code skipped bridge creation when `-offline` | Rebuild the `analysis` image from this branch (`make build/image/analysis`) — the fix is in `cmd/analyze/main.go` |
+| Dynamic offline run ends with `No matching distribution found for …` | Offline install can't reach the registry for dependencies | Expected for packages that need network at install time — run online, or analyze a package with vendored deps |
+| `Error: image '…' not found locally. Run 'make build' first.` (from the helper) | Images not built yet | Run the Step 1 build commands |
+| `Cannot connect to the Docker daemon` | Docker Desktop isn't running | Start Docker Desktop and wait until it reports "running" |
+| Mount / permission / file-sharing errors | Package path or `/tmp` not shared with Docker | Add the directory under Docker Desktop → Settings → Resources → File sharing |
+
+## See also
+
+- [README → Local Analysis](../README.md#local-analysis) — the condensed version
+- [sample_packages/README.md](../sample_packages/README.md) — about the test packages
+- [docs/data_schema.md](data_schema.md) — meaning of every result field
+- [docs/case_studies.md](case_studies.md) — real malware this project has caught
diff --git a/sandboxes/dynamicanalysis/Dockerfile b/sandboxes/dynamicanalysis/Dockerfile
index 2da8bf84..fb6c3c72 100644
--- a/sandboxes/dynamicanalysis/Dockerfile
+++ b/sandboxes/dynamicanalysis/Dockerfile
@@ -49,7 +49,6 @@ RUN apt-get update && apt-get -y upgrade && apt-get install -y --no-install-reco
 	libzip-dev \
 	net-tools \
 	netcat-openbsd \
-	powershell \
 	protobuf-compiler \
 	sshpass \
 	sudo \
@@ -61,13 +60,34 @@ RUN apt-get update && apt-get -y upgrade && apt-get install -y --no-install-reco
 	zip
 
 
+#
+# PowerShell setup
+#
+# Microsoft's apt repo (configured above) only publishes an amd64 PowerShell
+# .deb, so installing it via apt breaks the dependency solver on arm64. Install
+# from apt on amd64 (unchanged behaviour), and from the official linux-arm64
+# tarball build on arm64.
+ENV PWSH_ARM64_VERSION="7.6.3"
+RUN if [ "$(dpkg --print-architecture)" = "amd64" ]; then \
+		apt-get update && apt-get install -y --no-install-recommends powershell; \
+	else \
+		curl -fsSL "https://github.com/PowerShell/PowerShell/releases/download/v${PWSH_ARM64_VERSION}/powershell-${PWSH_ARM64_VERSION}-linux-arm64.tar.gz" -o /setup/powershell.tar.gz && \
+		mkdir -p /opt/microsoft/powershell/7 && \
+		tar zxf /setup/powershell.tar.gz -C /opt/microsoft/powershell/7 && \
+		chmod +x /opt/microsoft/powershell/7/pwsh && \
+		ln -sf /opt/microsoft/powershell/7/pwsh /usr/bin/pwsh; \
+	fi
+
+
 # Configure sudo for passwordless execution
 RUN echo "ALL ALL=(ALL) NOPASSWD: ALL" >> /etc/sudoers
 
 
 # install awscli v2
 # TODO: validate the integrity of the archive
-RUN curl "https://awscli.amazonaws.com/awscli-exe-linux-x86_64.zip" -o "/setup/awscliv2.zip"
+# AWS publishes per-arch installers; pick the right one so the image builds on
+# both amd64 (x86_64) and arm64 (aarch64) hosts.
+RUN curl "https://awscli.amazonaws.com/awscli-exe-linux-$(uname -m).zip" -o "/setup/awscliv2.zip"
 RUN unzip "/setup/awscliv2.zip" -d "/setup"
 RUN /setup/aws/install
 
diff --git a/scripts/load_sandbox_images_macos.sh b/scripts/load_sandbox_images_macos.sh
new file mode 100755
index 00000000..5bc29b3d
--- /dev/null
+++ b/scripts/load_sandbox_images_macos.sh
@@ -0,0 +1,86 @@
+#!/bin/bash
+#
+# Load locally-built sandbox images into a Docker named volume for use on
+# macOS (Apple Silicon or Intel).
+#
+# Background: the `analysis` image runs `podman` internally to launch the
+# dynamic/static sandbox containers. That nested podman reads its image store
+# from /var/lib/containers, which scripts/run_analysis.sh bind-mounts from the
+# host. On Linux dev boxes this is populated with `make sync` (sudo podman load
+# on the host). macOS has no host podman, and a plain host-directory bind mount
+# uses virtiofs, on which podman's overlay storage driver cannot create
+# whiteouts ("kernel does not support overlay fs").
+#
+# The fix is to back that store with a Docker *named volume* (stored on the
+# Docker Desktop VM's own filesystem, where overlay works) and to pre-load the
+# sandbox images into it once. run_analysis.sh treats a CONTAINER_DIR_OVERRIDE
+# value with no leading slash as a named volume, so:
+#
+#   export CONTAINER_DIR_OVERRIDE=pa-containers
+#   scripts/run_analysis.sh -nopull -local <pkg> ...
+#
+# will then find the images this script loads.
+#
+# Prerequisites: `make build` has produced the three local images
+# (analysis, dynamic-analysis, static-analysis).
+
+set -euo pipefail
+
+REGISTRY="gcr.io/ossf-malware-analysis"
+VOLUME="${CONTAINER_DIR_OVERRIDE:-pa-containers}"
+ANALYSIS_IMAGE="${REGISTRY}/analysis:latest"
+DYNAMIC_IMAGE="${REGISTRY}/dynamic-analysis:latest"
+STATIC_IMAGE="${REGISTRY}/static-analysis:latest"
+
+if ! docker info >/dev/null 2>&1; then
+	echo "Error: Docker is not running. Start Docker Desktop first." >&2
+	exit 1
+fi
+
+for img in "$ANALYSIS_IMAGE" "$DYNAMIC_IMAGE" "$STATIC_IMAGE"; do
+	if ! docker image inspect "$img" >/dev/null 2>&1; then
+		echo "Error: image '$img' not found locally. Run 'make build' first." >&2
+		exit 1
+	fi
+done
+
+# podman preserves the image config digest as its IMAGE ID, which matches
+# docker's image ID, so we can re-tag reliably by ID after loading.
+DYNAMIC_ID="$(docker image inspect "$DYNAMIC_IMAGE" --format '{{.Id}}' | cut -d: -f2 | cut -c1-12)"
+STATIC_ID="$(docker image inspect "$STATIC_IMAGE" --format '{{.Id}}' | cut -d: -f2 | cut -c1-12)"
+
+TMPDIR="$(mktemp -d)"
+trap 'rm -rf "$TMPDIR"' EXIT
+
+echo "Saving sandbox images (the dynamic image is several GB; this takes a while)..."
+docker save "$DYNAMIC_IMAGE" -o "$TMPDIR/dynamic.tar"
+docker save "$STATIC_IMAGE" -o "$TMPDIR/static.tar"
+
+echo "Creating named volume '$VOLUME' and loading images into the nested podman store..."
+docker volume create "$VOLUME" >/dev/null
+
+docker run --rm --privileged \
+	-v "$VOLUME:/var/lib/containers" \
+	-v "$TMPDIR/dynamic.tar:/dynamic.tar:ro" \
+	-v "$TMPDIR/static.tar:/static.tar:ro" \
+	-e "DYNAMIC_IMAGE=$DYNAMIC_IMAGE" \
+	-e "STATIC_IMAGE=$STATIC_IMAGE" \
+	-e "DYNAMIC_ID=$DYNAMIC_ID" \
+	-e "STATIC_ID=$STATIC_ID" \
+	"$ANALYSIS_IMAGE" \
+	bash -c '
+		set -e
+		podman load -i /dynamic.tar
+		podman load -i /static.tar
+		# podman load can lose the original repo tags; re-tag by image ID.
+		podman tag "$DYNAMIC_ID" "$DYNAMIC_IMAGE"
+		podman tag "$STATIC_ID" "$STATIC_IMAGE"
+		echo "--- sandbox images now available to podman ---"
+		podman images --format "{{.Repository}}:{{.Tag}} {{.ID}}"
+	'
+
+echo
+echo "Done. To run an analysis using these images:"
+echo
+echo "  export CONTAINER_DIR_OVERRIDE=$VOLUME"
+echo "  scripts/run_analysis.sh -nopull -local <package> -ecosystem <eco> -package <name>"