Logging Security

Rules (for AI agents)

ALWAYS

• Decide what is safe to log at the call site, and log a reference rather than a value: an id, a hash, a type, a count. A redactor is the last net, not the design — it matches field names and known patterns, so it cannot see personal data sitting inside an assembled string. A formatted notification body, a templated message, an exception with interpolated arguments: the name and the address are in the prose, in no named field, matching no pattern. Nothing downstream will catch that.
• Run a redactor at the sink anyway, for the cases the call site missed: credential-shaped values, Authorization and Cookie headers, query strings carrying tokens, and the personal-data patterns in rules/redaction_patterns.json. secret-detection owns which literals count as a credential; this rule is that the redactor exists and runs before anything leaves the process.
• Log in a structured format (JSON or logfmt) with stable field names — timestamp, service, version, level, trace_id, span_id, request_id, event, and user_id when there is an authenticated one. Structure is what makes redaction, retention, and access control expressible per field instead of per line.
• Neutralize user-controlled strings before they enter a log record. Line injection (CWE-117) is the classic case — a \n or \r lets an attacker forge a whole record — but structured logging moves the problem: a value that parses as JSON can inject fields into the ingested document, and ANSI escape sequences (\x1b[) are interpreted by the terminal of whoever reads the log. Encode the value; do not concatenate it into the line.
• Record security-relevant events as immutable audit records: authentication success and failure, MFA challenges, password and role changes, permission grants and revocations, data export, administrative action. These carry longer retention and stricter access than operational logs, and they are the reason failed logins must keep being written even when someone calls them noise.
• Treat read access to the log store as access to the data inside it. A centralized store aggregates every tenant's records, so grant it like production data access — named principals, reviewed, audited — not like an observability convenience handed to everyone who might want to debug something.
• Set retention per data category — short for debug, long for audit — and make it executed, not declared. A retention or deletion policy that no job enforces looks identical to one that works, right up to the point someone asks for evidence.
• Alert on the log pipeline itself: a service that stops logging, and volume that departs sharply from its own baseline in either direction. Calibrate the threshold against that service's normal traffic rather than a fixed multiple — a job that runs hourly and a request path have nothing in common.

NEVER

• Log a full request or response body, at any level. The level is not the control: a body captured at DEBUG in production is the same disclosure as one at INFO, and DEBUG in production is itself the finding.
• Log an entire bound SQL statement with its parameter values. Log the statement template, the parameter names, and a hashed identifier for the value.
• Use print() / console.log / fmt.Println in a production service. They bypass the redactor, the structure, and the sink — which is exactly the path by which the one unredacted value reaches disk.
• Write production logs only to local disk. They are lost with the pod, container or VM, which is usually the moment they were about to matter.

KNOWN FALSE POSITIVES

• Health-check and load-balancer probe logs downsampled or suppressed at the balancer to control volume.
• A request_id that happens to look like a token. Pattern redactors over-redact; allowlist your own correlation-id prefixes.
• Anonymous public-API access logs without auth headers are not in themselves a privacy problem — though client IPs may still be personal data under GDPR.
• Recording the type of record touched ("user accessed claim record") is the audit trail working. The rule is against field contents, not against recording that the access happened.
• Long audit retention, often years, is deliberate — an exemption from the general retention sweep rather than an oversight in it.

Context (for humans)

Logs are where secrets end up, and the reason is structural: logging is the one operation a developer performs without thinking about who reads the output. Request dumps, exception traces, debug prints and third-party SDK telemetry all default to "write everything", and the log store is then read by more people, for longer, than any database.

The order of the rules above is deliberate. Teams reach for a redactor first because it is a single place to fix everything, and it does catch the credential-shaped values. What it cannot catch is data that never looks like a credential — a customer name inside a rendered notification, an address in an exception message, a free-text field concatenated into an error. Those have no field name and match no pattern, and they are most of the personal data that leaks. The call site is the only place that knows what the value is.

The second thing worth stating plainly: the audit trail and the operational log are different products with different retention, different access, and different consumers. Storing them together is convenient and makes both harder to govern.

References

• references/verifying-findings.md — confirm or refute a finding, then lock it
• rules/redaction_patterns.json — the single redaction policy for this library; error-handling-security hands off to it rather than carrying its own
• rules/audit_event_schema.json
• OWASP Logging Cheat Sheet.
• CWE-532 · CWE-117.
• NIST SP 800-92.