October 2026 Security Audit
Every finding from the October 2026 audit is closed with a named regression test. The remediation is implemented and locally evidenced; it is not deployed and not certified. This page tells the story for operators: what was audited, what was found, what changed, and what the closure does and does not establish.
What the audit was
On October 23, 2026, hoike underwent a promise-to-code audit: a comparison of what the documentation promises (README, design document, ahu format specification, security and remediation documents, CLI help) against what the binaries actually do. It produced 12 findings: 5 High, 5 Medium, and 2 Low. Nine had isolated executable reproductions; the other three were source-traced. The audit baseline was commit c866700; the remediation merged in 41a534d.
This was the second review of the year. A September 2026 security review had already closed 29 findings (see the September remediation ledger); the October audit checked whether the remaining documented promises matched behavior.
Findings
| ID | Severity | Finding | Remediation |
|---|---|---|---|
| F01 | High | ahu verify reported success without verifying the CMS seal | Explicit verify modes: --config PATH (authenticated) or --structure-only; bare invocation fails. The admin verify route reports trust_status, trust_error, and signer_sha256, with overall_ok: false for untrusted input |
| F02 | High | Missing seal trust disabled authentication instead of failing admission | Fail-closed bundle admission: normal admission requires configured trust; only an explicit storage.allow_unauthenticated_bundles = true (with no other trust configured) permits structure-only demo admission |
| F03 | High | The serving loader accepted orphan deltas and ignored max_chain | Unsupported serving profiles (orphan deltas, delta chains, tombstones, compression) are rejected before any durable state change; storage.max_chain is removed, and configurations that set it fail to parse |
| F04 | High | Losing the state file silently reset anti-rollback protection | Strict state opening; hoike init-state bootstraps state once (no --force); serve never creates state implicitly; one writer per state directory, enforced with an OS flock held for the process lifetime |
| F05 | High | hoike import was an unauthenticated staging copy | hoike import --bundle PATH --ca LABEL --config PATH is an offline, authenticated, recoverable installation; --force is removed, and import fails while a server holds the writer lock |
| F06 | Medium | Zstd-marked payloads were served as raw compressed bytes | Compressed bundles are rejected at admission (the same gate as F03) |
| F07 | Medium | Seal/OCSP key separation was optional | Production file and PKCS#11 signing keys require a complete seal_key + seal_cert pair; ECDSA signers prove separation with a domain-separated signature verified under the seal public key; demo signing generates a distinct ephemeral seal pair |
| F08 | Medium | A PEM seal_cert failed sealing | Seal certificates are normalized (PEM or DER accepted), and the key/certificate pair is proven with a real CMS preflight seal on every load and signing pass |
| F09 | Medium | The signing-only gossip rollout dropped other upgraded peers | Explicit auth_mode policy (default required), peer_identities named-origin authorization, legacy peer_keys rejected with migration instructions, and a documented four-step fleet rollout |
| F10 | Medium | forward_ca was not wired into the TLS client | Per-CA forwarding trust roots are applied; without forward_ca, bundled WebPKI roots are used — never the operating-system trust store; unreadable or malformed roots fail startup and hoike check |
| F11 | Low | Gossip announcements carried the index digest | Announcements carry the manifest digest, so metadata-only generation changes are distinguishable |
| F12 | Low | The draft wire specification differed from produced bundles | The ahu 0.1 wire profile is frozen, with independently generated golden vectors in testdata/wire-profile-0.1/ |
Themes
Fail-closed bundle admission (F01, F02, F03, F06). Before the audit, several operator entry points treated a missing trust configuration as permission to skip authentication. Admission now requires configured seal trust everywhere: startup, hot reload, recovery, hoike check, and hoike import share one authenticated admission path, and unsupported serving profiles are rejected before any durable state changes. See Seal Trust Policy.
Durable-state integrity (F04, F05). Anti-rollback protection depended on a state file whose absence was treated as a fresh start. State is now bootstrapped explicitly with hoike init-state, opened strictly, guarded by an exclusive writer lock, and hoike import is an authenticated installation with an atomic commit point. See Anti-Rollback Protection.
Key separation (F07, F08). The seal key and the OCSP signing key must be distinct keys, configured as a complete pair, and proven to work together with a real CMS preflight seal before anything is published. Compromising or rotating one no longer implicitly affects the other.
Authenticated gossip (F09, F11). Gossip authentication is an explicit policy with a documented fleet rollout, and announcements identify a generation by its manifest digest rather than the index digest. See Gossip Configuration.
Trusted forwarding channels (F10). Each forwarding CA gets its own startup-validated outbound TLS client. A configured forward_ca becomes that CA’s only roots; a CA without one validates upstream against the bundled WebPKI roots, never the operating-system trust store.
Spec-truth (F12). The wire specification matches the bundles the tools produce, pinned by independently generated golden vectors. See The ahu Bundle Format.
Posture
All twelve findings are closed with named regression tests. The migration contract is executable and covered by subprocess A/B suites: a strict configuration with a persistent seal pair and existing state succeeds; new empty state fails until init-state; legacy max_chain and --force fail with actionable errors; peer_keys fails with migration instructions. The state-machine suites (state_lock, anti_rollback, installation) run on both Linux and macOS in CI.
The stated posture, verbatim: implemented and locally evidenced; not deployed, not certified. No production deployment, CC/FIPS validation, or independent security review follows from the closure. Live HSM, live 389 DS, real power-loss recovery, and multi-host qualification were not exercised.
October 2026 RustSec remediation
The same month, the dependency tree was remediated against the October 2026 RustSec advisory batch:
cargo auditreports zero vulnerabilities at the current lockfile and runs as a CI job onmainand on a monthly schedule. One non-fatal warning remains:rustls-pemfileis unmaintained.- The unused
zstddependency was removed. - The tree was consolidated on
p2560.14 andecdsa0.17, with a new internalecdsa_bridgethat keeps the wire encoding byte-identical. - TLS-stack updates were semver-compatible.
Evidence
- Promise-to-code audit, October 23, 2026 — the original findings, preserved unchanged as historical evidence
- Closure ledger — corrected behavior, responsible commits, regression tests, migration actions, and integrated release-evidence commands for each finding
- September 2026 remediation ledger — the 29-finding review that preceded the audit