View source for Postmortems/2026-05-24 Ultraplan Exposure

''As of: 2026-07-05 — migrated from `postmortems/2026-05-24-ultraplan-exposure.md`''

= Postmortem & Remediation Playbook: Ultraplan upload exposure (2026-05-24) =

'''Status:''' Remediation in progress. This document is BOTH the postmortem
and the step-by-step playbook for executing the remediation manually.

== Incident summary ==

While working on the tsbuffer project plan in Claude Code, the <code>/ultraplan</code>
command was invoked. <code>/ultraplan</code> uploads the current working directory
plus <code>CLAUDE.md</code> and <code>MEMORY.md</code> to an Anthropic-hosted "Claude Code on
the web" session for remote planning by an agent there.

The remote session terminated with <code>error_during_execution</code> before
producing any output. The uploaded content remains in Anthropic cloud
session storage tied to session ID <code>01GUjtsXpaAw1CqHShfePPmq</code>.

'''No warning was surfaced about the upload before it occurred.''' This is
the root cause of the exposure — there is no broken security control to
fix; the tool worked exactly as designed, and we did not know what it did.

== What was in the upload ==

=== Live credentials (rotation required) ===

{| class="wikitable"
|-
! Item
! Where it was
! Why it's bad
|-
| n8n API JWT
| <code>CLAUDE.md</code>, "n8n API" section, pasted in plaintext, no expiry
| Anyone with this can call any n8n API endpoint as admin
|-
| n8n admin UI password (<code>Htdml0fI1aZgtPBVv5oz</code>)
| <code>MEMORY.md</code>, n8n DB Recovery memory entry
| Direct UI login as <code>admin@wilsoz.com</code>
|}

=== Topology / metadata exposed (no rotation possible, hygiene only) ===

* Internal domains: <code>agent.wilsoz.com</code>, <code>vault.wilsoz.com</code>, <code>key.wilsoz.com</code>,
  <code>photos.wilsoz.com</code>, <code>ntfy.wilsoz.com</code>, <code>mymail.wilsoz.com</code>,
  <code>next.wilsoz.com</code>, <code>wilsoz.cloudflareaccess.com</code>.
* Internal IPs: <code>192.168.0.109</code> (stronghold), <code>192.168.0.144</code> (ollama),
  Docker bridges, Tailscale IP <code>100.67.169.50</code> for the SDF relay VPS.
* OpenBao endpoint and secret path layout (paths only, no tokens).
* Vaultwarden item-naming convention ("Stronghold Infrastructure" folder).
* MXRoute server <code>blizzard.mxrouting.net</code>, username <code>craniums</code>.
* Authentik admin URL + admin usernames <code>akadmin</code>, <code>marcus</code>; Cloudflare
  Access OIDC client ID.
* Matrix bot @zordon:wilsoz.com, room ID, internal Synapse endpoint.
* All n8n workflow JSONs in <code>n8n-workflows/</code> and every service doc in
  <code>services/</code>.
* Email addresses including <code>marcus@wilsoz.com</code>, <code>admin@wilsoz.com</code>,
  <code>jkatz@craniumslows.com</code>.

== Top-level constraints for this remediation ==

# '''The credentials store IS the root of trust.''' Vaultwarden + OpenBao
   are how we manage every other secret. Compromising either means
   compromising everything else.
# '''Never paste a secret into chat, ever.''' Not even temporarily, not
   even "just to verify." Once it's in the conversation, it's leaked
   exactly the same way the original JWT was.
# '''Verify each new credential works BEFORE deleting the old one.''' A
   half-rotated credential where the new one doesn't work is worse than
   the original exposure.
# '''No secret values in command-line arguments.''' They show up in
   <code>/proc/PID/cmdline</code> and <code>ps auxw</code> output. Use stdin, env vars, or
   token files mode 0600.
# '''Two-person rule (where applicable).''' For high-blast-radius
   rotations (root tokens, master password) — don't rotate without
   confirming the recovery path first.

----

== Remediation steps (manual playbook) ==

Each step has: (a) what to do, (b) how to verify, (c) what to update,
(d) rollback.

=== Step 1. Rotate n8n API JWT ===

'''(a) Action:'''
* Open https://agent.wilsoz.com in a browser; log in as <code>admin@wilsoz.com</code>.
  Password: retrieve from Vaultwarden item "n8n UI Login". (This is the
  old password — Step 2 will rotate it. The order matters: API key first,
  because once we change the admin password we want to verify with the
  API that everything's healthy.)
* Settings → n8n API → see the existing key (do NOT copy its value).
* Click "Revoke" on the existing key. Confirm.
* Click "Create an API key". Set "Time to live" to "No expiration".
  Copy the new JWT once shown (the UI shows it exactly once).
* '''Paste the new JWT directly into a Vaultwarden item''' — open
  Vaultwarden in another tab, edit item "n8n API Key", paste, save. Do
  not paste it into a terminal, a text file, or any chat.

'''(b) Verify:'''
<syntaxhighlight lang="bash">
# Retrieve the new key from Vaultwarden via the bw CLI (NOT by pasting):
NEW_KEY=$(bw get item "n8n API Key" | jq -r '.fields[] | select(.name=="apiKey") | .value')
# (Field name may differ — adjust per your existing Vaultwarden item layout.)
curl -sf "https://agent.wilsoz.com/api/v1/workflows" \
  -H "X-N8N-API-KEY: $NEW_KEY" \
  | jq '.data | length'
# Expect: a number (count of workflows), no errors.
unset NEW_KEY
</syntaxhighlight>

'''(c) Update:'''
* Vaultwarden item "n8n API Key": done above.
* OpenBao at <code>secret/n8n</code> field <code>api-key</code>:
  ```bash
  # Read the new key from Vaultwarden into a shell var (no echo):
  NEW_KEY=$(bw get item "n8n API Key" | jq -r '.fields[] | select(.name=="apiKey") | .value')
  # Write to OpenBao using stdin (NOT command-line arg):
  printf '%s' "$NEW_KEY" | BAO_TOKEN="$(cat /library/openbao-app/.claude-mcp-token)" \
    bao kv put -mount=secret n8n api-key=-
  unset NEW_KEY
  ```
* <code>~/Developer/SystemResiliency/CLAUDE.md</code>: replace the JWT-bearing line
  with: <code>Retrieve from Vaultwarden item: "n8n API Key"</code>. Remove the
  <code>N8N_KEY='...'</code> example line that contains the actual JWT in the
  curl examples; replace with a <code># N8N_KEY=$(bw get item "n8n API Key" | ...)</code>
  pattern.
* Commit CLAUDE.md change. Verify the diff contains NO JWT value before
  committing (<code>git diff CLAUDE.md</code> and visually scan).

'''(d) Rollback:'''
If verification fails, the old JWT was already revoked at n8n — there is
no rollback to the previous value. Instead, immediately mint another new
JWT from the n8n UI and retry. The window of "no working API key" is
acceptable because no automation actively uses it during this rotation.

=== Step 2. Rotate n8n admin UI password ===

'''(a) Action:'''
* Still logged into n8n as <code>admin@wilsoz.com</code>.
* Settings → Personal → Change Password.
* Open Vaultwarden, edit item "n8n UI Login", click the password
  regenerate button (32 chars, full charset). Save.
* Copy the new password from Vaultwarden.
* Paste into n8n's "New Password" field. Save.

'''(b) Verify:'''
* Sign out of n8n.
* Sign in with email <code>admin@wilsoz.com</code> and the new password.
* Expect: dashboard loads, workflows visible.

'''(c) Update:'''
* Vaultwarden item "n8n UI Login": done above.
* OpenBao at <code>secret/n8n</code> field <code>admin-password</code> (if used by anything):
  check first whether anything reads it. If yes, update via stdin pattern
  above. If no, skip.
* <code>MEMORY.md</code> (auto-memory): edit <code>/home/mnw/.claude/projects/-home-mnw-Developer-SystemResiliency/memory/MEMORY.md</code>.
  Find the line: <code>Created new owner: admin@wilsoz.com / Htdml0fI1aZgtPBVv5oz</code>
  Replace with: <code>Created new owner: admin@wilsoz.com / (rotated 2026-05-24; see Vaultwarden "n8n UI Login")</code>

'''(d) Rollback:'''
Same situation as Step 1 — old password is already invalidated. If the
new password doesn't work, n8n's password reset relies on SMTP email
which is functional. Worst case: re-init the n8n DB (same recovery we
did 2026-03-11).

=== Step 3. Audit CLAUDE.md and MEMORY.md for OTHER secrets ===

This is a one-pass sweep to catch anything I missed when triaging.

'''(a) Action:'''
<syntaxhighlight lang="bash">
cd ~/Developer/SystemResiliency
# Install gitleaks if not present:
which gitleaks || sudo apt install -y gitleaks
# Scan the project + memory dirs:
gitleaks detect --source . --no-git -v
gitleaks detect --source ~/.claude/projects/-home-mnw-Developer-SystemResiliency/memory --no-git -v
</syntaxhighlight>

'''(b) Verify:'''
* Read the findings line by line.
* For each finding: is it actually a secret, or a false positive (e.g.
  an ID, a hash of a non-sensitive value)?
* True positives: add to the rotation list, restart from Step 1 for each.

'''(c) Update:'''
* Document any new findings in this file (append a "Found in audit"
  subsection below).

'''(d) Rollback:'''
N/A — read-only operation.

=== Step 4. Attempt deletion of the cloud session ===

'''(a) Action:'''
* Open https://claude.ai/code/session_01GUjtsXpaAw1CqHShfePPmq in a
  browser. Log in as <code>marcus@wilsoz.com</code>.
* Look for a delete / remove option in the session menu (three-dot icon,
  settings gear, or similar).
* If found, click it. Confirm.
* If not found: send an email to <code>privacy@anthropic.com</code> with the session
  ID and a request for deletion under data-deletion rights. Reference
  https://privacy.anthropic.com. Subject:
  <code>Data deletion request — session 01GUjtsXpaAw1CqHShfePPmq</code>.

'''(b) Verify:'''
* If UI delete worked: refresh the session URL, expect 404 or
  "session not found".
* If email request: save the sent message; expect a reply within 30 days
  per their stated SLA.

'''(c) Update:'''
* Append the outcome to this file: "Cloud session deletion: [UI delete |
  email requested on YYYY-MM-DD | response received on YYYY-MM-DD]"

'''(d) Rollback:'''
N/A. Best-effort operation either way.

=== Step 5. Install a pre-commit secret-scanner ===

So a future paste-in is caught before it reaches a commit.

'''(a) Action:'''
<syntaxhighlight lang="bash">
cd ~/Developer/SystemResiliency
# Add a .gitleaks.toml if you want custom rules; defaults are fine.
# Install the pre-commit framework if not already:
which pre-commit || sudo apt install -y pre-commit
cat > .pre-commit-config.yaml <<'EOF'
repos:
  - repo: https://github.com/gitleaks/gitleaks
    rev: v8.18.0
    hooks:
      - id: gitleaks
EOF
pre-commit install
</syntaxhighlight>

Do the same in <code>~/Developer/jellyfin-adhoc/</code> and <code>~/Developer/tsbuffer/</code>.

'''(b) Verify:'''
* Make a trivial change to a file in the repo.
* Stage it. Try to commit.
* Expect: pre-commit runs, gitleaks scans the staged diff, commit
  proceeds (because the diff is clean).
* Manually test the catch: stage a file containing <code>EXAMPLE_KEY=eyJabc.def.ghi</code>,
  commit; expect gitleaks to block.

'''(c) Update:'''
* Commit the <code>.pre-commit-config.yaml</code> to each repo.

'''(d) Rollback:'''
<code>pre-commit uninstall</code>.

=== Step 6. Adopt a policy on cloud-handoff tools ===

<code>/ultraplan</code> is convenient but its blast radius is the entire working
directory. Several reasonable options:

* '''Block it entirely''' via a Claude Code hook in <code>~/.claude/settings.json</code>.
* '''Allow but require confirmation''' via a pre-tool hook that shows the
  user the upload manifest before proceeding.
* '''Sanitize first''' — never invoke from a directory that contains
  CLAUDE.md / MEMORY.md.

Decision pending — capture in <code>~/Developer/SystemResiliency/TODO.md</code> for
follow-up after this incident closes.

----

== Outcome (filled in as work happens) ==

* [ ] Step 1 — n8n API JWT rotated
* [ ] Step 2 — n8n admin password rotated
* [ ] Step 3 — Audit complete; findings:
* [ ] Step 4 — Cloud session deletion attempt:
* [ ] Step 5 — Pre-commit hook installed
* [ ] Step 6 — Cloud-handoff policy decision:

== Follow-ups (for TODO.md) ==

* Cloud-handoff tool policy decision (Step 6 above).
* After this exposure, audit existing OpenBao <code>scripts-n8n</code> and
  <code>claude-mcp</code> tokens: are they scoped tightly enough? Should they be
  rotated on a schedule (e.g. every 90 days) regardless of leaks?
* Consider whether Vaultwarden master password should be rotated on a
  cadence (high blast radius if leaked; can't be auto-rotated).

== Lessons ==

# Any Claude Code tool that "uploads context" needs an explicit
   user-confirmation prompt with the manifest of what will be sent. The
   silent default is a footgun.
# CLAUDE.md and MEMORY.md are read by everything: every Claude Code
   session, plus any tool that exports project context. They must never
   contain literal secrets. The bar should be "would I publish this to
   a public README?" — if no, it doesn't belong in those files.
# Rotation tedium is the actual problem behind "I'll just inline this
   credential for now." The exposure-remediation script (see
   <code>scripts/exposure-remediation/PLAN.md</code>) is the load-bearing follow-up.