ObliviOS Developer portal Project site

Home / Decisions / ADR-113

Retire every stallion key generated on 2026-09-28 as exposed, and generate a fresh stallion Personal Stable key set before any Personal Stable ceremony.

Accepted, recorded 28 September 2026.

Decision
ADR-113
Date
2026-09-28
Status
AcceptedAccepted
Mentions
ADR-111

1Decision

Retire every stallion key generated on 2026-09-28 as exposed, and generate a fresh stallion Personal Stable key set before any Personal Stable ceremony. This supersedes ADR-111 decision (2), the re-designation of that set. Two rules are also enforced on the signing host for every device: a ceremony refuses to run unless <vault> is an actual mountpoint, and it refuses to run while any non-zram swap is active. Memory shortfalls are fixed with more signing-host RAM (owner target 6GB, up from 1.9GB).

2Why and evidence

Owner asked for a new key; read-only inspection of the signing host the same day (agent access granted by the owner) confirmed it is required, not just preferable. (a) With the vault closed, the plain directory under the mountpoint (<vault>/keys/stallion/ with the vault unmounted, on the unencrypted /home LV) held a complete private key set in plaintext: all nine .pk8, avb.pem, id_ed25519, dated 09:28.

These were writes made while the vault mount had silently failed. The in-vault avb.pem is the same AVB key (SHA-1 089b874ee4419a8c39c37374386a161a6cf2ee26), so that key has sat on unencrypted disk. (b) /proc/swaps showed disk-backed swap (/dev/dm-2, the swap_1 LV) active by default on every boot, including during that day's attempts, one of which was OOM-killed in re-verification.

(c) One attempt (09:23) completed generate-release.sh, so this key signed a userdebug OTA. (d) The scripts' vault check tested only that releasekey.pk8 exists, which the plaintext copy satisfied. With the vault closed, a ceremony would have signed from the unencrypted copy. (e) Every failed attempt's generate-release.log lived in /dev/shm and was wiped by the shutdown trap. That is why seven failures that day had no readable cause.

3Consequences

Owner actions, all on the signing host: destroy the plaintext copies under the vault mountpoint with shred -u (best effort only on a journaled filesystem on LVM, which is exactly why the keys are retired rather than cleaned and reused); delete the retired set inside the vault; swapoff the swap_1 LV, remove it from /etc/fstab and zero it; raise the VM's RAM.

Then generate a new set with the vault mount verified (mountpoint) immediately before and after, .pk8 via make_key with a blank password and avb.pem via openssl genrsa 4096 \| openssl pkcs8 -topk8 -nocrypt. The latter is required by decrypt-keys' -n $password check (.claude/local-server-memory.md). Record the new AVB fingerprint.

tools/signing/sign-ota.sh and sign-baseline-images.sh now enforce both host rules in a preflight that runs before any staging dir or shutdown trap exists (verified on the host: vault closed → VAULT NOT MOUNTED ... refusing to run, exit 1, VM stays up), and keep the last 300 lines of generate-release.log/verify.log in ceremony-log/ on failure.

Not decided here: the lynx Alpha key went through the 2026-09-27 ceremony with disk swap deliberately re-enabled, the same exposure class as (b). Whether to rotate it is an open owner decision.

4Review when

New key generated and fingerprint recorded; any future finding of key material outside the mounted vault; any ceremony that needs disk swap to finish.

Read from docs/governance/decision-log.md, entry ADR-113.