Sota Secrets Management

SOTA Secrets Management

Purpose

Eliminate static secrets where possible; where not possible, make every secret short-lived, narrowly scoped, runtime-injected, auditable, and rotatable without downtime. This skill covers the full lifecycle (generation → distribution → storage → use → rotation → revocation → expiry), storage backends, application handling patterns, leak detection, incident remediation, and per-credential-type rules. It serves two workflows: BUILD (write correct secrets handling into new or existing code) and AUDIT (sweep a repo for secret issues and report findings).

The hierarchy of preference, always:

1. No secret at all — workload identity / OIDC federation / cloud IAM roles.
2. Short-lived, auto-issued secret — Vault dynamic creds, STS tokens, SPIRE SVIDs.
3. Long-lived secret in a managed backend — secret manager + rotation + audit log.
4. Encrypted secret in the repo — SOPS+age / sealed-secrets, GitOps only.
5. Plaintext secret anywhere — never acceptable.

When you write or review code, push the design as far up this hierarchy as the platform allows, and document why if you stop below level 2.

BUILD mode

Use when implementing anything that consumes or manages a credential.

1. Classify the credential. Type (DB cred, API key, signing key, TLS key, …) determines the rules — read rules/05-credential-types.md for the matching section before writing code.
2. Try to eliminate it. Cloud-to-cloud or CI-to-cloud calls should use workload identity (OIDC, IAM roles, SPIFFE) — see rules/01-lifecycle-and-workload-identity.md. Only fall back to a stored secret when no federation path exists (e.g., third-party SaaS API key).
3. Pick the storage backend per environment using the decision table in rules/02-storage-backends.md. Never invent a custom encrypted store.
4. Wire injection at runtime — file mount or env var populated by the platform, fetched via SDK with caching/TTL, never baked into images or code. Patterns and good/bad pairs in rules/03-application-patterns.md.
5. Design rotation before shipping. Every secret needs: an owner, a rotation procedure that works with zero downtime (dual-secret / kid overlap), an expiry or rotation interval, and a revocation path. If you cannot answer "how do we rotate this at 3am during an incident," the design is not done.
6. Add guardrails: pre-commit scanning config, CI secret-scan gate, .gitignore entries for .env* and key files, log-scrubbing for the new secret's shape (rules/04-detection-and-remediation.md).
7. Self-review against the Audit checklist at the end of every rules file you used.

AUDIT mode

Use when asked to find secret leaks/misuse in an existing repo.

Sweep procedure

1. Tooling pass (if available): run gitleaks git --redact . and/or trufflehog filesystem . (and git log history scan when the repo has history). Treat tool output as candidates, not verdicts — verify each hit.
2. Manual grep pass for what tools miss. Sweep at minimum:
   • High-entropy strings and known prefixes: AKIA, ASIA, ghp_, gho_, github_pat_, xoxb-, xoxp-, sk-, sk_live_, rk_live_, AIza, ya29., glpat-, npm_, dop_v1_, shpat_, eyJhbGciOi (inline JWTs), -----BEGIN .* PRIVATE KEY-----.
   • Assignment patterns: (password|passwd|pwd|secret|token|api[_-]?key|auth)\s*[:=]\s*['"][^'"]{6,}.
   • Connection strings with embedded creds: ://[^/:@\s]+:[^@\s]+@ (postgres, mysql, mongodb, amqp, redis URLs).
   • Files: .env* tracked in git, *.pem, *.p12, *.pfx, *.key, *.jks, *.keystore, id_rsa*, credentials.json, serviceaccount*.json, kubeconfig, .npmrc/.pypirc with tokens, terraform.tfstate (state files contain plaintext secrets).
3. History pass: git log -p / gitleaks git --log-opts for secrets removed from HEAD but live in history. A secret deleted in a later commit is still leaked — severity is unchanged.
4. Handling pass (misuse, not just leaks): secrets in log statements, error messages, exception payloads, crash/telemetry dumps, URLs/query strings, CLI args (visible in ps), Dockerfile ENV/ARG, docker-compose environment: literals, Kubernetes manifests with stringData/base64 secrets committed, CI YAML with inline values, debug endpoints dumping config, world-readable key files, missing rotation/expiry on long-lived tokens, overly broad token scopes.
5. Verify and triage each finding: is the value real (test-shaped? placeholder? entropy?), is it currently valid, what blast radius. Never call a credential live by invoking it against production without explicit permission; judge from context.

Severity conventions

Confirmed-fake placeholders (changeme, xxx, <YOUR_KEY>, obvious test fixtures) are not findings, but note them as Low if they are realistic enough to mask real leaks in scans.

Finding format

Report every finding as:

[SEVERITY] path/to/file.py:123 — RULE-ID short title
  Evidence: the offending line, with the secret value REDACTED (show prefix + length only)
  Why: one sentence of impact
  Fix: concrete remediation (rotate first, then remove; target pattern to adopt)
  Effort: trivial | small | medium | large

Order findings Critical → Low. End the audit with: counts per severity, whether git history is affected (if so, remediation must include rotation + history purge per rules/04-detection-and-remediation.md), and the top 3 systemic fixes. Never reproduce a discovered secret in full in your report — redact to first 4 chars + length.

Rules index

Top-10 non-negotiables

1. A leaked secret is rotated first, scrubbed second. History rewriting without rotation is theater — assume every secret that ever touched git, logs, or a ticket is compromised.
2. No secrets in source code or VCS, ever — including "temporarily," tests against real services, example files with live values, and committed .env files.
3. Prefer no secret to a managed secret: if OIDC federation / IAM roles / SPIFFE can replace a static credential (CI→cloud, service→cloud, pod→service), use it. Static keys for cloud access from CI are a defect, not a choice.
4. Every secret has an expiry or rotation interval and a documented zero-downtime rotation procedure (dual-secret overlap, JWT kid, DB dual users). Unrotatable = misdesigned.
5. Generate secrets with a CSPRNG, ≥256 bits of entropy for opaque tokens (≥32 random bytes before encoding); never derive from timestamps, UUIDv4-as-secret, or human-chosen strings.
6. Secrets never appear in: logs, error messages, exception payloads, crash dumps, URLs or query strings, CLI arguments, ps output, Dockerfile layers, image env, shell history, or telemetry. Scrub at the logger and wrap in redacting types.
7. Inject at runtime, never at build time. Images, artifacts, and bundles are secret-free; the platform (orchestrator, secret-manager SDK, CSI driver) supplies values when the process starts — file mounts preferred over env vars where supported.
8. Least privilege and per-environment separation: one credential per consumer per environment, scoped to the minimum actions/resources; dev/staging/prod never share secrets and prod values are unreadable from non-prod.
9. Application data encryption uses KMS envelope encryption (encrypt data with a data key, wrap the data key with a KMS key); never hardcode or hand-manage raw encryption keys.
10. Scanning is mandatory and layered: pre-commit (gitleaks/trufflehog) on every developer machine, a blocking CI gate, and periodic full-history scans. Detection without a leak runbook is incomplete — keep the rotate→revoke→purge→monitor runbook current.