@syncrona/redaction

| | | | | | |:--:|:--:|:--:|:--:|:--:|

Shared secret detection and redaction for SyncroNow AI. This package is the single source of truth for the question "is this a credential?", asked by every component that persists ServiceNow data or tool traffic:

• the MCP server audit trail — packages/mcp-server/src/audit.ts
• the instance mirror redactor (@syncrona/mirror), which writes an entire instance into a git tree

The detectors grew inside the audit trail, hardened one security review at a time. They moved here when the mirror needed them, because two copies of a secret scanner drift, and the copy nobody is actively reviewing is the one that falls behind — in the consumer that writes to a repository rather than a local log.

What lives here

Sensitive key names

• isSensitiveKey(key) — true when a property NAME suggests credential material (password, apiKey, x-auth-token, session, passphrase, otp, …). Deliberately over-eager: it also matches keyword and monkey. A redacted key name costs a reviewer one lookup; a leaked credential costs a rotation.

Secret-shaped values

• looksLikeSecretValue(value) — true when a VALUE looks like credential material regardless of the key it arrived under: scheme://user:pass@host and bare user:pass@host connection strings, JWTs, PEM private keys, inline Authorization headers, AWS access-key IDs, vendor-prefixed API keys (Stripe, OpenAI, GitHub, Slack, GitLab, Google) and raw 256-bit hex key material.
• SCAN_BUDGET (8192) — the maximum length scanned in full. Fails closed: a longer value is reported as a secret without being scanned. The historical behaviour was the opposite, and padding a secret past the budget was all it took to launder it into a log in cleartext.

Stable redaction markers

• redactValue(value) — __SYNCRONA_REDACTED__<sha256-12>.
• REDACTION_MARKER_PREFIX, REDACTION_MARKER_HASH_CHARS.

The marker embeds a short digest of the plaintext, so the same secret yields the same marker on every sync (a mirror re-run of an unchanged instance must be byte-identical) while a rotated secret yields a different one. A constant marker would erase that history: a credential could be replaced with no diff at all.

Design notes

This module is intentionally a leaf: it depends on no other @syncrona package, only on node:crypto. A security primitive at the bottom of the dependency graph can be audited on its own and can never be made to import something that imports it back.

It is also pure and synchronous — no I/O, no clock, no configuration. The MCP server calls it while holding the audit lock and the mirror calls it once per field of every record on an instance, so the cost must be bounded and the result must be a function of the input alone.

What is not here

Field-type deny — dropping a column because sys_dictionary reports its internal_type as password/password2 — belongs to the caller, which has the schema knowledge this package deliberately does not. It is nevertheless the strongest of the three signals: measured on a real instance, a password2 column returns a 106-character ciphertext over the Table API, which no value-shape rule here would flag. Never weaken that path on the strength of this one.

False positives are a cost, not a bonus

A match redacts the whole value, so an over-eager value pattern destroys the forensic detail of the record an operator is reading. Two patterns here have already been narrowed after exactly that regression (a bare Basic/Bearer keyword matched ordinary prose; a bare user:pass@host form matched npm:[email protected]). Each carries its honest limit in a comment beside it. The corpus tests in test/ assert both directions — every rule has a true positive and a near-miss that must pass through verbatim.