Threat model
Assets
Section titled “Assets”The roster (who may approve what), device keys, the release and catalog signing keys, the core’s CA and audit key, the secrets the core seals (provider tokens, forge tokens), the audit chain, and the systems Vaicoa changes.
What it is designed to withstand
Section titled “What it is designed to withstand”| Threat | What Vaicoa does |
|---|---|
| An agent that is wrong or manipulated (prompt injection from logs, tickets, web pages) | it can only propose catalog actions in its mayPropose list; policy decides; risky actions need a human; its text is shown as an unverified claim |
| A proposer who dresses a dangerous action in a harmless description | the display is rendered from the signed payload by the signed template; claims come last and are never the basis of the decision |
| A compromised core trying to run something on a node | the node verifies human approvals against the roster it pinned, the pinned catalog and the plan; it refuses older rosters and replayed playbook steps |
| A forged or replayed approval | signatures over a canonical digest (RFC 8785) that binds payload, template, evidence and confirmation; expiry; idempotency keys |
| State that drifts between approval and execution | the plan is bound; drift voids the approval |
| A lost or stolen phone | suspension by another device of the member or by an owner, at once; its pending decisions are dropped |
| A leaked session ticket | single use, at most ten minutes, bound to the session’s OIDC token or ServiceAccount |
| Tool poisoning through MCP | tool names and descriptions come only from the signed catalog; imported MCP tools are pinned by a digest of their definition and blocked on drift |
| Rewriting history | the hash-chained audit with a signed head; exports carry their chain proof |
What it does not withstand, or not yet
Section titled “What it does not withstand, or not yet”- A compromised core can withhold, delay or refuse work, and can refuse to sign its audit head.
- Agent approvals are not verified by nodes. A compromised core could claim a reviewer agent approved a
routineaction. Risky actions are not affected: they always need a human, which the node verifies. - Third-party API actions (Hetzner Cloud, Forgejo, Alertmanager, observability, MCP servers) run in drivers at the core; for them the core is the only checkpoint.
- Display templates are not checked by nodes against their pinned templates (release builds close this through pinned release key ids).
- A stolen audit key can sign a shortened head. Write the head down elsewhere regularly.
- A fully compromised owner device together with its user verification is out of scope.
- Phones, YubiKeys and the Mac app are unproven on real hardware (limitations).
- Supply chain: there is no release key yet, no second signature on it, and no TUF.
- No independent security audit.
Reviews so far
Section titled “Reviews so far”The code has had two internal code reviews and an automated multi-agent review of the fixes. All of them were
done by AI agents (Claude); no person outside the project and no security firm has audited the code yet. Every
finding was reproduced with a failing test and fixed. The largest, a review of the trust path (node verification, the core’s approval path, the
phone routes, crypto and supply chain, outputs and drivers), found 33 issues: 2 critical, 2 high, 12 medium and 17
low. The automated review of those fixes found 10 more, among them one new hole the fixes had opened. All were fixed on main by 7 November 2026. This is not a
security audit, and Vaicoa makes no claim to be secure.
Found something? Report it privately to security@mitjamartini.com.