This document describes what the project is for, what it is not for, and which risks remain in the code as it stands.

A local network you control, where every device either is yours or has been allowed on purpose:

• home LAN;
• isolated office or lab network;
• a bench of your own devices.

• The open internet without a VPN or tunnel on top. Discovery is broadcast UDP and assumes a flat, trusted segment.
• Networks where untrusted third-party devices roam. Anyone on such a segment can see discovery traffic and attempt to connect.
• Anything with a compliance requirement. This code has not been audited.

A single root CA signs every node certificate. Nodes accept a neighbour only if its certificate chains to that CA and is not revoked. Discovery packets are signed with the Ed25519 key stored in the sender's certificate, but a packet from an unknown node is treated as a connection request, not as proof of identity — real authentication happens in the mTLS handshake. Whoever controls the CA can therefore mint any identity in the network. The CA key is the crown jewel: it is stored on disk, optionally encrypted with AES-256-GCM, and not protected by an HSM.

1. Revocation depends on connectivity. CRLs propagate over gossip. A node isolated from the network does not learn it was revoked until it reconnects. The local cache survives restarts but cannot help an isolated node.
2. SSH gives a shell as the beacon's user. Typically root. The guard is the allowlist: SSH is off unless -ssh is given, and an empty allowlist denies everyone but the node itself. Anyone who obtains a server-role certificate and is listed can run commands as that user.
3. No rate limiting on SSH authentication. There is a connection-count limit (-ssh-max-conns), but no delay between failed attempts.
4. Time-based. Discovery packets are accepted within ±5 minutes. Nodes with badly drifting clocks cannot see each other.
5. No key rotation. Certificates expire (90 days by default) and can be reissued, but there is no automatic rotation and no pinning of the CA certificate beyond the local copy.
6. window-change is not forwarded to the interactive process, so terminal resizing does not reach the shell.
7. Written by AI agents. The code is reviewed by a human, not by a cryptographic audit. Assume the usual "good for a well-understood network" bar, not more.

Open an issue. For anything that looks like a vulnerability, please describe reproduction steps rather than a theoretical concern — this is a small project and concrete reports get fixed.