Fail-closed by construction.

packages/tenancy-rls's buildTenantPolicySql emits ENABLE ROW LEVEL SECURITY plus FORCE ROW LEVEL SECURITY on every tenant table, so the policy binds the table owner too, not just other roles. The sole entry point, withTenant, opens a transaction, drops to the unprivileged app role, and binds the account id into a Postgres GUC (app.current_account) that every policy reads; a code path that forgets withTenant has no GUC bound and the table returns nothing. A one-time role pre-flight (assertRoleNotPrivileged) refuses to run if that role is ever a superuser or BYPASSRLS, since either would silently no-op FORCE ROW LEVEL SECURITY. Cross-tenant isolation is a CI test, not a convention.

packages/kernel's ssrf.ts stops DNS rebinding on every user- or config-supplied URL (the alerting webhook transports and the AI-Production provider baseUrl both route through it. assertSafePublicUrl rejects non-https, credentials-in-URL, and a literal private/loopback/link-local/metadata host at the config boundary; assertResolvedHostPublic then resolves the hostname and re-checks every returned A/AAAA record against the same denylist immediately before the outbound fetch, so a public name that DNS-rebinds to 127.0.0.1 or 169.254.169.254 is caught where a literal-only check can't see it. ssrfGuardedFetch forces redirect: "error") only the original host is re-checked, so a followed redirect could otherwise carry the request past the guard.

Evidence buckets enable S3 Object Lock. The default is GOVERNANCE mode, inside the retention window an object cannot be overwritten or deleted by an app bug or an ordinary operator, though a caller holding s3:BypassGovernanceRetention can still override it. COMPLIANCE mode is available as an explicit, irreversible opt-in (production-only, gated behind irreversibleComplianceOptIn) for retention even the AWS account root cannot shorten.

Each audit row commits SHA-256 over the previous hash plus its own payload. Tampering with any historical row breaks every link after it, and the break is detectable, provable, and exportable.

packages/kernel's crypto.ts is the one home for secret comparison: safeEqualFixed converts both sides to equal-length buffers and runs node:crypto's timingSafeEqual for session tokens and HMAC outputs of known length; safeEqualVariable SHA-256-hashes both sides first for variable-length values like an admin-email allowlist entry, because a raw variable-length timingSafeEqual throws on a length mismatch and leaks a boolean through the catch. verifyAllowlisted scans every allowlist entry with no early return, so the timing never reveals which entry matched.

Sensitive columns are sealed with a data key derived per tenant via HKDF-SHA256. A leaked tenant key exposes one tenant, never the table; rotating the root re-derives every key without a re-encrypt scan.