
Give each MCP tool its own scope, and return a refusal the client can act on
Part 6 gated the whole server with one scope, which meant the token that let an agent look up a customer also let it issue refunds. Splitting that per tool is easy. Making the refusal useful is not: the obvious implementation returns HTTP 200 with the error buried in the body, so a client has nothing to trigger step-up authorization on. Getting the 403 and WWW-Authenticate challenge the spec defines means leaving the framework's abstraction entirely — and the version that does it decodes the token a second time, unverified.

Sandbox the code your agent writes, and prove every limit actually applied
OWASP names running model-generated code through exec or eval as its own vulnerability, tells you to treat the model as untrusted — and says nothing about how to contain it. So we built the containment: no network, 256 MB, 64 processes, read-only disk, no capabilities. Every one of those held. The CPU limit was refused outright, and eight busy loops took 605% of the host. The cause is one line of systemd configuration nobody mentions, and the fix needed no reboot despite the documentation insisting otherwise.

Move the Docker daemon off root so a container escape lands on an ordinary user
If you can run docker without sudo, you can already read every file on the machine. Not through a bug — that is how Docker is built. One command proves it. For a host running agents that execute generated code, that is the whole attack, and Part 2's non-root containers do not touch it. This moves the daemon itself to an unprivileged account, with the UID arithmetic you can check, four places the install walks you into a wall on Ubuntu, and what stops working afterwards.

Firewall an agent container so it can reach one API and nothing else
Every part of this series so far has controlled what an agent may call. None of it controls where an agent may go. OWASP's mitigations for Excessive Agency are three rules about tools and say nothing about the network. This puts a default-deny egress policy in front of one container, allows exactly one API, and proves the boundary with a lookup that succeeds and a packet that dies. Includes the conntrack rule everyone tells you to add and this one doesn't need, and a measurement that was wrong by three orders of magnitude.

The wrong valid token: authenticating an MCP tools server with authentik
Agent orchestration part 4 scoped tools in the client and admitted the server still trusted whoever reached the port. This puts a JWT verifier in front of it, with its own authentik client, so the tools server refuses anonymous callers — and refuses the gateway's own valid token, because a separate provider means a separate issuer. Includes the documented client-credentials helper that cannot work here, and the four silent 404s it produces instead of saying so.

The key that expires: giving an agent its own identity at the gateway
Part 4 ended with a shell, a static key and a model that answered. This replaces that key with a token your identity provider issues, that carries a scope, and that dies after five minutes. authentik issues it over client credentials, the gateway verifies it in twenty-five lines, and three status codes prove the boundary. Includes the requirement authentik doesn't document, and the licence wall you hit if you follow LiteLLM's own guide.

Split one MCP toolbox between two agents so the scheduler cannot issue refunds
Part 3 handed one agent all five tools and let a human pause guard the dangerous one. That guard depends on the model asking. This splits the same MCP server between a scheduler and a billing agent, so the scheduler cannot refund an invoice for the simplest reason available — the tool was never in its list. Then the supervisor goes back on top, and routing turns out not to be the guarantee.

The binding that wasn't there: group access control, and what SSO doesn't protect
Part 3 shipped SSO with empty bindings. Closing that gap on the LiteLLM gateway turned up something worse: authentik's wizard let us configure the bindings, showed them in its own table, and saved none of them — no error, no warning. Then, once access control actually worked, the API answered a request from a shell with no account, no session and no membership. Both of those are the post.

One login for everything: putting authentik in front of a self-hosted app
Part 2 fixed how containers run. This fixes who gets to log in. Deploy authentik as your own identity provider, connect Langfuse to it over OIDC, and understand the pieces — provider, application, redirect URI, scopes — instead of copying a config. Includes the two failures a real run produced: a compose file that silently swallows your settings, and OAuthAccountNotLinked.

Root by default: hardening container privilege on a self-hosted AI stack
Two of the six containers behind a live Langfuse deployment were running as root, for no reason anyone chose. Fixing it took one line each — and broke a service that had nothing to do with the fix. A real conversion, a real coordination failure, and how to catch both.

Your firewall is lying to you: hardening Docker networks for multi-agent systems
A box running five agent stacks, a firewall set to deny everything, and services still answering from the public internet. We probe a real deployment, find a database with no password, prove why UFW never sees Docker traffic, and fix it four ways — every command and result from a live run.

Make OpenClaw private with a NetBird mesh VPN
Put OpenClaw on a private mesh with NetBird, repoint your hostname at the mesh IP, and close the public ports — so the same chat URL works only for your devices. Real commands and gotchas from a live run.

Secure OpenClaw with a Caddy reverse proxy + HTTPS
Put Caddy in front of OpenClaw for real HTTPS and device-paired auth, then close the gateway's ports so the proxy is the only way in — gotchas and all.