Skip to content
A personal experiment · alpha · code not public yet · no offer, no support Project status →

Security model

The idea in three sentences: humans prove themselves with a key in hardware, and how much that proof counts is its level. An agent only supplies data: it proposes, the policy decides as for a human, and risky actions always need a second human at level presence. Whoever approves sees a display their own client renders from signed data, and exactly that display is signed.

An agent run canAn agent run cannot
propose the actions in its trigger’s mayPropose list, from the signed catalogpropose anything outside that list or the catalog
ask explain what the core would decideapprove its own proposal, or a risky action at all
read metrics, logs, traces and Kubernetes objects within a budget, redactedread secrets; read in production unless you enabled agent-read there
act with at most the rights of the user it acts for, capped at devicereach presence, or exceed its token, proposal and time budget
write text that the approver sees as an unverified claimchange the facts, the plan or the display the approver signs

A reviewer agent can settle safe and routine actions under the starter and standard profiles, never risky ones, and only advises under strict.

  • Per action: the catalog’s parameter schema and CEL constraints bound every parameter (for example a volume can only grow, within the environment’s ceiling).
  • Per environment: roles can be scoped to environments, and policy decides per environment what runs on its own.
  • Per run: budgets for tokens, runs per day, proposals per run and time.
  • Per playbook: maxShare caps the share of a cluster touched at once.
  • Per change window: freezes and spent error budgets stop changes.
  • Destructive actions need the target picked or typed, are never retried, and never go as an emergency change.

For actions that a node runs (Podman, host, k3s, Kubernetes, Cluster API), the node verifies the op itself: the signature against the roster it pinned (it refuses an older roster), the pinned catalog and risk class, the plan, the frames in the roster, and every playbook step against the approved expansion. A compromised core can withhold or delay work there, but it cannot make a node run a risky op no human approved.

For actions against third-party APIs (Hetzner Cloud, Forgejo, Alertmanager, observability, MCP servers) a driver runs the op at the core, and the core is the checkpoint. The node cannot verify agent approvals yet. Both are listed in the threat model.

One setting picks a consistent set of defaults. The crypto profile is the same in all three; profiles change how much proof is demanded, never which algorithms are used.

starter (default of init)standardstrict
Device level at pairingthe client’s claim countsunattested keys capped at device; presence needs attestationevery device attested
Risky quorumowner or operator at presence, one approver ≠ proposersametwo approvers ≠ proposer, distinct persons
Agent approversmay settle safe/routinemay settle safe/routineadvisory only
Approval lifetime24 h1 h15 min
Standard changes, solo modeallowed, shown as relaxedaudited weakeningsrefused

Under strict an explicit setting may only tighten; the core refuses to start otherwise. Raising the profile re-evaluates every paired device, and vaicoa-core check lists who can no longer approve what before the change takes effect. The profile is signed into the roster, so nodes enforce it too.

Vaicoa follows BSI TR-02102: TLS 1.3 only with AES-GCM suites and the groups P-384 and P-256, ECDSA P-256/P-384 keys, SHA-256/384, AES-256-GCM, HKDF-SHA-256; IPsec (IKEv2) instead of WireGuard. Not used: Ed25519, X25519, ChaCha20-Poly1305, WireGuard, age. Every connection between two Vaicoa components uses this profile without exception and refuses any RSA key in the peer’s chain.

Two documented deviations:

  • RSA ≥ 2048 bits for third-party TLS peers only (container registries, object stores, MCP servers, push services, the release feed’s web host, identity providers, webhooks). TR-02102-2 asks for 3000 bits. A test fails if this configuration appears anywhere else. Content fetched over such a connection is verified by signature or digest.
  • TLS 1.2 for the S3 client of the core’s backup.

Keys can live in PKCS#11 (HSMs, YubiKeys), the Secure Enclave and the Android Keystore. The core’s IKE key is a file for now.

Every release is to carry SHA256SUMS, its ECDSA P-256 signature and CycloneDX SBOMs per artefact. The built-in catalog is signed with the release key and pinned by the core and the nodes. Updates come through a signed release feed with an anti-rollback check; each node verifies an update against its own pinned release key, and the binary that installs an update is never the one being updated. The release key has not been created yet.

Privately by email to security@mitjamartini.com (security.txt), not in a public issue. Acknowledgement within 3 working days, a first assessment within 10, coordinated disclosure with credit; details and the safe-harbour terms are in the repository’s SECURITY.md.