diff --git a/SECURITY.md b/SECURITY.md
new file mode 100644
index 0000000..77642b8
--- /dev/null
+++ b/SECURITY.md
@@ -0,0 +1,82 @@
+# Security Policy
+
+## Supported versions
+
+Only the most recent release is supported. Fixes ship as a new release rather than as patches
+backported to older tags — see [`CHANGELOG.md`](CHANGELOG.md) for what has landed.
+
+If you consume this template as a git submodule, pin it to a **tag** rather than a bare commit.
+A bare commit from a branch that is later squash-merged becomes unreachable, and every fresh clone
+of your repo then fails its `git submodule update`.
+
+## Reporting a vulnerability
+
+Report privately through GitHub: **[Security → Report a
+vulnerability](https://github.com/TheHefty/jvsl.env.agents.code-server/security/advisories/new)**.
+
+Please do not open a public issue for something you believe is exploitable.
+
+This is a personal project with a single maintainer, so there is no guaranteed response time and no
+bounty. Expect a best-effort acknowledgement, and please include what you did, what happened, and
+what you expected instead — a reproduction against a container built from `setup` is worth far more
+than a description of a suspicious-looking flag.
+
+## What this template is, and what it is not
+
+This is a **single-user development environment** for a personal workstation. It is not a
+multi-tenant sandbox, not a hosting platform, and not a boundary you should place between yourself
+and code you actively distrust.
+
+The boundary that is meant to hold is the **host boundary**: nothing running inside the container —
+including the Claude Code agent — should be able to reach the host's filesystem, its Docker daemon,
+or its network beyond what was deliberately handed in. The host's Docker socket is
+deliberately *not* mounted for exactly this reason; `docker` inside the container talks to a nested
+rootless daemon that cannot see the host's containers. If you find a way across that line, that is a
+vulnerability and we want to hear about it.
+
+Boundaries *inside* the container are weaker on purpose, and are documented as such below.
+
+### In scope
+
+- Escaping the container to the host — filesystem, daemon, or privileged host resources.
+- The nested rootless daemon reaching the host's daemon, containers, or images.
+- code-server's port becoming reachable beyond the host's loopback interface (it is published as
+  `-p 127.0.0.1:0:8443`).
+- Credentials or tokens baked into the built image, or leaked from the bind-mounted
+  `/config` / `~/.claude` state to somewhere they should not be.
+- Supply-chain problems in the build itself: a `Dockerfile.frag` fetching an artifact over an
+  unverified channel, a compromised or typosquatted dependency, a pinned digest that does not
+  match what it claims.
+
+### Out of scope — deliberate design decisions
+
+These are known, documented, and accepted. Reports about them will be closed as intended behaviour,
+though a report arguing the *tradeoff itself* is wrong is a reasonable thing to open a normal issue
+about.
+
+- **`ai-jail`'s restrictions are advisory within the container.** They bound the agent's own shell,
+  not everything the agent can reach. In particular, an agent granted `--rw-map /config/.docker`
+  can talk to the nested daemon, and `docker run -v /:/probe` against that daemon hands it this
+  container's own root filesystem — including paths the sandbox hides. Writes are bounded by the
+  rootless uid mapping (container-root maps to `abc`), not by `ai-jail`. This was measured rather
+  than assumed, and the socket was kept knowingly; the reasoning is in
+  [`docs/OVERVIEW.md`](docs/OVERVIEW.md) under "Why the container is this permissive".
+- **code-server runs with no password.** `start` passes an empty `PASSWORD=`, so any user or process
+  on the host that can reach the published loopback port gets the editor, and through it a shell in
+  the container. This is a single-user-workstation assumption, not an oversight.
+- **`--cap-add=SYS_ADMIN` and `--security-opt seccomp=unconfined` / `systempaths=unconfined`.**
+  These exist so `ai-jail`'s `bwrap` sandbox can create user namespaces on distros whose AppArmor
+  policy restricts unprivileged namespace creation. Without them `ai-jail` cannot build its sandbox
+  at all. The properly narrow fix is a host AppArmor profile, which a container cannot install for
+  itself.
+- **Nested containers get no cgroup limits of their own.** `/sys/fs/cgroup` is read-only, so the
+  rootless daemon cannot enforce per-container cpu/memory. Everything stays bounded by the outer
+  container's limits, which is the containment that matters here.
+- **Anything you deliberately grant.** `--device` passthrough, extra `--rw-map` entries in a
+  project's `.ai-jail`, and mounts you add yourself widen the boundary by design.
+
+## Reviewing this yourself
+
+`docs/OVERVIEW.md` records the reasoning behind every permissive flag, including the decisions that
+were reversed and why. If you are evaluating whether to adopt this template, that document — not
+this file — is the honest account.