docs: activate credential break-glass planning
Assistant: codex Assistant-Model: gpt-5.6-sol Assistant-Session: 01a0290b-3241-74c3-b868-6049545af836
This commit is contained in:
parent
5c2368ad42
commit
fff76ef089
5 changed files with 169 additions and 23 deletions
117
docs/credential-governance-break-glass.md
Normal file
117
docs/credential-governance-break-glass.md
Normal file
|
|
@ -0,0 +1,117 @@
|
|||
# Credential governance break-glass contract
|
||||
|
||||
Status: active design contract for WARDEN-WP-0027-T02. This document does not
|
||||
authorize a seal, re-key, token mint, policy write, workload restart, or host
|
||||
reboot.
|
||||
|
||||
## Decision boundary
|
||||
|
||||
ops-warden consumes and verifies credential-control evidence; it does not own
|
||||
the OpenBao cluster, unseal shares, recovery snapshots, policies, or root token.
|
||||
`railiance-platform` is accountable for the OpenBao operation. The share
|
||||
custodians, platform driver, abort operator, provider-console operator, and
|
||||
affected workload owners participate through an attended approval window.
|
||||
|
||||
The current production trust-root is the rotated Shamir barrier restored to
|
||||
railiance01 in RMASTER-WP-0020: three separately custodied shares with a
|
||||
threshold of two. Share-holder identities and share material remain in the
|
||||
approved out-of-band custody system, never Git, State Hub, shell history, logs,
|
||||
or chat. A readiness receipt names participating *roles* and attests that two
|
||||
distinct custodians are present; it never contains a share or recovery value.
|
||||
|
||||
Root is offline bootstrap/break-glass material only. It is not a substitute for
|
||||
`platform-admin` OIDC and must not be used through the browser UI. An ordinary
|
||||
unseal uses threshold shares and does not require a root token.
|
||||
|
||||
## Graded response
|
||||
|
||||
| Grade | Trigger | Action owner | Exit evidence |
|
||||
| --- | --- | --- | --- |
|
||||
| Observe | Suspicion without confirmed credential disclosure | lane owner + risk-nexus | Metadata-only taint, audit, policy, and capability review |
|
||||
| Soft lockdown | Coding-agent access must stop while human recovery remains available | railiance-platform | `agent-high-risk-boundary` source/live equality; all concrete high-risk data paths denied; metadata readable |
|
||||
| Lane containment | One or more concrete credentials are exposed | each credential owner | Front door disabled, provider/OpenBao rotation through the owner-native procedure, consumer cutover, taint cleared only after verification |
|
||||
| Hard lockdown | OpenBao trust or control-plane integrity is in doubt | platform driver under attended approval | Fresh encrypted snapshot receipt, intentional seal, sealed-state proof, 2-of-3 unseal, post-unseal and consumer verification |
|
||||
| Re-key | A Shamir share or the barrier custody model is compromised | platform owner + threshold custodians | Separately approved OpenBao re-key ceremony, new threshold custody attestation, old-share retirement, recovery drill |
|
||||
|
||||
Hard lockdown is not an agent command. The irreversible hold point is the
|
||||
platform owner's attended seal action. All prerequisites below must be true
|
||||
before that hold point; repository access alone grants no authority to cross it.
|
||||
|
||||
## Pre-seal hold point
|
||||
|
||||
- An approved, bounded window names the accountable platform driver and a
|
||||
distinct abort operator.
|
||||
- Two distinct Shamir custodians attest availability through the approved
|
||||
out-of-band channel.
|
||||
- Independent provider-console access is verified by its owner.
|
||||
- A current Raft snapshot is encrypted and stored off-host; the non-secret
|
||||
receipt binds cluster id, Raft index, plaintext/encrypted hashes, location,
|
||||
age, and verification without including protected material.
|
||||
- Current OpenBao health, seal state, Raft peers, audit device, auth methods,
|
||||
policy fingerprints, and consumer readiness are captured as metadata.
|
||||
- A rollback/re-entry order and stop conditions are acknowledged by affected
|
||||
workload owners.
|
||||
|
||||
Any missing or stale item aborts before sealing. A live incident may require
|
||||
immediate network isolation, but that does not authorize improvising share or
|
||||
root-token handling.
|
||||
|
||||
## Re-entry sequence
|
||||
|
||||
The platform owner performs the exact commands from
|
||||
`railiance-platform/docs/openbao.md`; ops-warden does not copy a second
|
||||
execution procedure here.
|
||||
|
||||
1. Confirm the instance is intentionally sealed and the approved window is
|
||||
still open.
|
||||
2. Two custodians supply shares through hidden, non-logged prompts. No agent
|
||||
observes the values.
|
||||
3. Prove `initialized=true` and `sealed=false`, then run the owner post-unseal
|
||||
verification.
|
||||
4. Verify Raft, persistent audit output, OIDC metadata, SSH roles, and policy
|
||||
fingerprints before restoring normal access.
|
||||
5. Reconcile ExternalSecret stores and bounded consumers in dependency order.
|
||||
6. Run value-safe capability probes, including the coding-agent deny-wins check
|
||||
and an ops-warden signing smoke that records only backend and decision ids.
|
||||
7. Revoke temporary operator tokens and close the window with non-secret
|
||||
timing, status, and abort/rollback evidence.
|
||||
|
||||
The August 3 migration already proved two attended restart/unseal cycles under
|
||||
the same 2-of-3 barrier. T02 still requires one current emergency-seal drill
|
||||
against the authoritative runtime because restore rehearsal and intentional
|
||||
production lockdown are different claims.
|
||||
|
||||
## Warden signing recovery
|
||||
|
||||
Do not apply the parked standalone `warden-sign` AppRole as a T02 break-glass
|
||||
path.
|
||||
|
||||
- Normal operation uses the railiance-platform credential broker and short-lived
|
||||
`warden-sign` child token.
|
||||
- Broker/issuer recovery uses an attended OIDC platform operation.
|
||||
- A sealed or integrity-compromised OpenBao is recovered by the Shamir
|
||||
trust-root; another AppRole cannot bypass the seal.
|
||||
- Placing standing AppRole material outside the broker expands credential
|
||||
custody without improving root recovery.
|
||||
|
||||
The AppRole dry-run in SECRETS-WP-0004 remains useful negative evidence: its
|
||||
capability set is narrow. It is not apply authorization. A separately approved
|
||||
ops-bridge unattended-signing design may re-evaluate it under its own workplan;
|
||||
that service-access question is outside T02 and does not create recovery
|
||||
authority.
|
||||
|
||||
## Evidence and owner surfaces
|
||||
|
||||
- Current barrier/re-entry evidence: `railiance-master` RMASTER-WP-0020 and
|
||||
`docs/evidence/openbao-isolated-restore-2026-08-03.json`
|
||||
- Authoritative execution and recovery runbook:
|
||||
`railiance-platform/docs/openbao.md`
|
||||
- Current coordinated recovery gates:
|
||||
`railiance-platform/docs/railiance01-coordinated-reboot.md`
|
||||
- Emergency evidence contract and validator:
|
||||
`railiance-platform/docs/openbao-emergency-drill-evidence.example.json` and
|
||||
`make openbao-validate-emergency-evidence`
|
||||
- Soft-lockdown policy and proof: railiance-platform RAILIANCE-WP-0022 and
|
||||
ops-warden `scripts/check_agent_read_boundary.py`
|
||||
- Warden-sign recovery input: secrets-engine SECRETS-WP-0004 and
|
||||
`workplans/ADHOC-2026-08-11.md` T03
|
||||
Loading…
Add table
Add a link
Reference in a new issue