Promote railiance-backup-offsite-lane to active/resolvable after capabilities-safe re-verify. Add catalog risk=high, agent read-boundary (exit 7 + OpenBao policy companion), EXPOSED taint via warden taint, and close WP-0026.
12 KiB
| id | type | title | domain | repo | status | owner | topic_slug | planning_priority | planning_order | created | updated | state_hub_workstream_id |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| WARDEN-WP-0026 | workplan | Credential disclosure hygiene + rotation guidance (Strand A) | infotech | ops-warden | finished | codex | custodian | high | 26 | 2026-07-16 | 2026-07-16 | 1adb34af-d5f8-42c5-89b4-44593995a281 |
Credential disclosure hygiene + rotation guidance (Strand A)
Origin
Follow-up to a credential-disclosure incident on 2026-07-16 while verifying
CCR-2026-0004 (railiance offsite backup lane). A negative policy test was run
as BAO_TOKEN=$(bao token create -policy=default -field=token) bao kv get <path>.
The token create was denied (workload role lacks it), BAO_TOKEN fell back to
the caller's privileged login token, and bao kv get printed all three field
values (NC_WEBDAV_TOKEN, NC_WEBDAV_URL, AGE_PRIVATE_KEY) into an agent
session transcript. Buildup mode — exposure accepted, learnings captured.
Root causes: (1) a deny-test that read the secret data path at all; (2) a silent privileged-token fallback; (3) the read landed in a logged context.
This is Strand A (disclosure hygiene). Tamper-resistant policy governance and one-command mass rotation/lockdown (Strand B) are deliberately deferred — see "Out of scope" below.
Goal
Make accidental secret disclosure structurally hard, and make ops-warden the authoritative source for how each secret is rotated or re-established — advisory knowledge held next to the routing catalog, not in OpenBao.
Design guardrails (binding on acceptance)
- Verification never reads secret data. Use
bao token capabilities(allow/deny) instead ofkv getfor positive/negative lane tests. - Masking is defense-in-depth, not a boundary. Any display filter is a wrapper
convenience; raw
bao kv get <path>remains the documented anti-pattern. - Rotation guidance covers both
rotate(provider re-mint) andre-establish(regenerate from source, e.g. a new age keypair). - Coverage gate: every
active(and newly promoted) catalog lane must carry rotation guidance, enforced by a scorecard check.
Out of scope (Strand B — deferred, not built here)
- Executable one-command rotation of all tainted secrets.
- Graded lockdown / break-glass seal + re-key with a designed trust-root.
- Policy time-travel / policy-as-code reconcile-to-past-commit. (Drift detection may be revisited separately; the reconcile machinery is out.)
Task: Capabilities-based lane verification
id: WARDEN-WP-0026-T01
status: done
priority: high
state_hub_task_id: "9329e72d-c07c-41ce-88ce-e8602eb72c43"
Done 2026-07-16: canonical capabilities-safe verification pattern added to
wiki/playbooks/catalog-lane-promotion.md (fleet promotion checklist criterion 8
- dedicated section), and applied to the
railiance-backup-offsite-laneandforgejo-admin-api-tokenplaybook verify sections. Positive/negative proven viabao token capabilitiesagainst the KV v2 data path — neverbao kv get; the denieddefaulttoken-create is documented as a pass, not a fallback trigger. Live CCR-2026-0004 re-verify carried under T07.
Replace secret-reading verify flows with capability checks. Positive test:
approved identity has read on the KV data path. Negative test: a default-only
identity is deny. Both via bao token capabilities <token> <path> (or the
self endpoint) — never bao kv get. Update
wiki/playbooks/*lane*.md verify sections and any lane-verify helper.
Done when: the documented and tooled verify path for any lane proves allow/deny without reading a secret value, and the CCR-2026-0004 re-verify uses it (see T07).
Task: Safe access transport (no stdout values)
id: WARDEN-WP-0026-T02
status: done
priority: high
state_hub_task_id: "3f28c573-8e58-4851-8aa0-925f9367f266"
Done 2026-07-16: sanctioned transports added to warden access so a value never
lands on stdout — --out FILE (mode-0600 file), --exec (child env, pre-existing),
and --wrap (single-use OpenBao response-wrapping token via bao kv get -wrap-ttl,
caller bao unwraps in their own context). Raw --fetch to a non-TTY stdout is now
refused (exit 6) unless --unsafe-stdout is passed (interactive human only).
proxy_fetch_to_file/proxy_fetch_wrapped/build_wrapped_fetch in proxy.py; tests
in tests/test_proxy.py. Anti-pattern + transports documented fleet-wide in
.claude/rules/credential-routing.md and wiki/OperatorAccessAssist.md (G2).
warden access / fetch paths must emit values only into an env var, a file, or
a response-wrapping token (bao … -wrap-ttl), never a stdout table. Add a
wrapping-token transport for values that must move between processes. Record in
canon (credential-routing rules) that raw bao kv get <path> (full table) is
the anti-pattern; the sanctioned path is warden access … --field <F> into env.
Done when: the sanctioned fetch path cannot print a value to a terminal, and the anti-pattern is documented fleet-wide.
Task: Masking display filter (defense-in-depth)
id: WARDEN-WP-0026-T03
status: done
priority: medium
state_hub_task_id: "21ab08d5-7782-4567-a08d-980211dd7851"
Done 2026-07-16: warden/mask.py (fingerprint/mask_value — presence, length,
8-char sha256 prefix; never the value) + proxy_fetch_fingerprint and a
warden access … --fingerprint masked status view (bypasses the stdout guard
because it emits no value). Lets two parties compare sha256 prefixes to confirm a
shared value (e.g. rotation landed) without disclosure. Explicitly labelled
defense-in-depth — raw bao kv get bypasses it — in wiki/OperatorAccessAssist.md
and the module docstring. Tests in tests/test_mask.py + a CLI test in
tests/test_proxy.py.
In the warden wrapper, mask KV data values by default when any listing/status is shown — display presence, length, and a short non-reversible hash instead of the value. Explicitly labelled as defense-in-depth (raw bao bypasses it).
Done when: wrapper-mediated output never shows a raw KV value, with the limitation documented.
Task: Agent read-boundary on high-risk lanes
id: WARDEN-WP-0026-T04
status: done
priority: high
state_hub_task_id: "f95d4381-f995-4c9d-b438-70b4f0ed90c5"
Done 2026-07-16: Catalog risk: high|standard (default standard). High-risk:
railiance-backup-offsite-lane, forgejo-admin-api-token, openrouter-llm-connect.
OpenBao policy agent-high-risk-boundary (railiance-platform + live write) grants
metadata/capabilities only and denies data-read on those paths — verified with
minted agent token (data=deny, metadata=read). warden access with
WARDEN_AGENT_ID set refuses raw value stream on high-risk lanes (exit 7);
--out/--exec/--wrap/--fingerprint remain. Playbook:
wiki/playbooks/agent-read-boundary.md. Tests in tests/test_routing.py +
tests/test_proxy.py.
Repo: railiance-platform (OpenBao policy/roles). Agent identities receive
capabilities/metadata and wrapping tokens on high-risk lanes, not raw data
reads. Align with the existing credential-routing rule ("ops-warden proxies reads
as the caller and must not retain values"). Classify which lanes are high-risk
(recovery escrow like AGE_PRIVATE_KEY, upload tokens).
Done when: at least the high-risk lanes deny raw data reads to agent roles while still allowing wrapped/proxied access, verified via capabilities checks.
Task: EXPOSED taint convention
id: WARDEN-WP-0026-T05
status: done
priority: medium
state_hub_task_id: "e989f8f0-930e-4d2f-9037-221e04f34199"
Done 2026-07-16: Convention documented in wiki/playbooks/exposed-taint.md
(exposed_at, exposed_version, exposed_reason, exposed_ref on KV v2
custom_metadata). First worked mark applied to
platform/workloads/railiance/backup/offsite-lane version 2 (disclosure
incident). warden taint <id> (+ --json) reports taint via metadata-only
bao kv metadata get — never secret data (src/warden/taint.py). Tests in
tests/test_taint.py.
Repo: railiance-platform (OpenBao) + ops-warden surface. Establish a KV v2
custom_metadata convention to mark a tainted secret: exposed_at=<datetime>
and the affected version. Identify semi-automatic candidates from the OpenBao
audit log (reads from agent/shared contexts). warden surfaces taint status for
a lane (advisory; no auto-rotation here).
Done when: a secret can be marked EXPOSED via a documented convention and
warden reports whether a lane is currently tainted.
Task: Rotation / re-establishment guidance registry
id: WARDEN-WP-0026-T06
status: done
priority: high
state_hub_task_id: "b8a43aee-18c7-46cd-9cfe-1f73234746d3"
Done 2026-07-16: rotation: block (method rotate|re-establish, ordered steps,
owner, automatable) added to the routing model/parser (RotationGuide,
RouteEntry.rotation, vends_secret), screened for secret material in a prose-safe
mode. warden rotate-guide <id> (human + --json) surfaces the guidance; warden route show --json carries has_rotation + rotation. Coverage enforced by the new
catalog_rotation_coverage scorecard check (every active secret-vending lane must
have a block) and promotion checklist criterion 9. Rotation blocks authored for all
7 active vending lanes + the draft railiance-backup lane (re-establish example: age
keypair regen + re-encrypt). Tests in tests/test_routing.py. Also fixed a
pre-existing keyword collision (bare npm on the forgejo-admin lane → forgejo-npm)
so "npm token" routes to the generic lane again.
Give every catalog lane structured-but-advisory renewal guidance, held in the
ops-warden registry (not in OpenBao). Add a rotation: block per catalog entry
capturing: method (rotate | re-establish), ordered steps (provider re-mint /
keygen / OpenBao write / re-encrypt-and-reupload where relevant), owner, and
automatable (bool, for future Strand-B). Surface via warden rotate-guide <id>
(and/or warden route … --rotate). Add a scorecard coverage check: every
active lane must have a rotation: block; flag any that don't.
Done when: warden rotate-guide <id> returns actionable renewal steps for
every active lane, and the scorecard fails if any active lane lacks guidance.
Task: Incident lessons + first worked lane (CCR-2026-0004)
id: WARDEN-WP-0026-T07
status: done
priority: medium
state_hub_task_id: "9944f46d-3706-43a4-9300-7f63bf87c9ff"
Done 2026-07-16: Lessons note present. Capabilities-safe live re-verify on
bao.coulomb.social: lane-policy token → read on data path; default + agent
boundary → deny; field keys present (no values printed). Catalog promoted
draft→active, fetch_command pinned to NC_WEBDAV_TOKEN (no placeholders)
so resolvable: true; risk: high; rotation guidance + EXPOSED taint on v2.
CCR-2026-0004 evidence + access_frontdoor.resolvable: true / readiness: ready.
Playbook + wiki/CredentialRouting.md updated. Operator may still rotate the
exposed values optionally (buildup) — not a promotion blocker.
Write a short lessons-learned note (buildup context; exposure accepted; the three
root causes). Apply T01 + T06 to CCR-2026-0004 as the first worked lane:
re-verify it the capabilities-safe way so it can finally promote to
resolvable: true (unblocking RAILIANCE-WP-0015), and ensure its rotation:
block (rotate Nextcloud token; re-establish age keypair + re-encrypt artifacts)
is present. Rotation of the exposed values is the operator's optional call, not
a blocker (buildup).
Done when: the lessons note exists, CCR-2026-0004 has capabilities-based verify + rotation guidance, and its promotion path is unblocked.
References
CCR-2026-0004-railiance-backup-offsite-lane.yaml(railiance-platform)wiki/playbooks/railiance-backup-offsite-lane.md.claude/rules/credential-routing.md(the-custodian, fleet-inlined)RAILIANCE-WP-0015(railiance-apps) — cnpg backup coverage, gated on CCR-2026-0004- Strand B (deferred): tamper-resistant governance, one-command rotation/lockdown, policy time-travel — capture separately if/when justified.