diff --git a/audit-ci.json b/audit-ci.json
index c9e616cf..8282639b 100644
--- a/audit-ci.json
+++ b/audit-ci.json
@@ -1,8 +1,11 @@
 {
   "package-manager": "pnpm",
-  "high": true,
   "critical": true,
+  "high": true,
+  "moderate": true,
   "report": true,
   "summary": true,
-  "allowlist": ["GHSA-86w9-cpqp-85rv"]
+  "allowlist": [
+    "GHSA-86w9-cpqp-85rv"
+  ]
 }
diff --git a/backend/server/createApiServer.ts b/backend/server/createApiServer.ts
index 0647b49c..783255da 100644
--- a/backend/server/createApiServer.ts
+++ b/backend/server/createApiServer.ts
@@ -18,6 +18,8 @@ import { randomBytes } from 'node:crypto';
 
 import express, { type Express, type Request, type Response } from 'express';
 import { cacheHeadersMiddleware } from '../shared/middleware';
+import { createCspMiddleware, DEFAULT_CSP_POLICY } from '../services/shared/cspMiddleware';
+import { createCsrfMiddleware } from '../services/shared/csrfService';
 import {
   createPublicApiRouter,
   createSearchRouter,
@@ -160,6 +162,17 @@ export function createApiServer(options: CreateApiServerOptions = {}): Express {
 
   app.disable('x-powered-by');
 
+  // ── Security headers (CSP + HSTS + framing) — #1769 ──────────────────────
+  // Applied before all routes so every response carries the policy.
+  app.use(
+    createCspMiddleware({
+      ...DEFAULT_CSP_POLICY,
+      // Collect violation reports at a dedicated endpoint so we can monitor
+      // policy regressions in production without blocking legitimate traffic.
+      reportUri: ['/api/v1/csp-report'],
+    }) as unknown as express.RequestHandler,
+  );
+
   // Request id, structured access log and server span for every request.
   app.use(
     createRequestObservability({
@@ -195,6 +208,42 @@ export function createApiServer(options: CreateApiServerOptions = {}): Express {
   // body into memory before it is parsed.
   app.use(express.json({ limit: MAX_JSON_BODY }));
 
+  // ── CSRF protection — #1770 ───────────────────────────────────────────────
+  // Double-submit cookie strategy: safe methods (GET/HEAD/OPTIONS) issue a
+  // token; unsafe methods verify that X-CSRF-Token matches __Host-csrf.
+  // Webhook paths are skipped — they use HMAC-based authentication instead.
+  app.use(
+    createCsrfMiddleware({
+      cookie: { sameSite: 'Strict', secure: process.env['NODE_ENV'] === 'production' },
+      skipPaths: [
+        '/api/v1/stripe/billing/webhook',
+        '/api/v1/paddle/webhook',
+        '/api/v1/webhooks',
+      ],
+    }) as unknown as express.RequestHandler,
+  );
+
+  // Bootstrap endpoint: clients call GET /api/v1/csrf-token once on startup
+  // to receive their first token before issuing any mutating request.
+  app.get('/api/v1/csrf-token', (_req: Request, res: Response) => {
+    // The CSRF middleware already set the cookie + header on this safe GET.
+    res.json({ ok: true });
+  });
+
+  // CSP violation report collector — browsers POST here when a directive is
+  // violated.  The body is logged at warn level and discarded (no state is
+  // mutated, so it is exempt from CSRF verification by the skip-path above).
+  app.post(
+    '/api/v1/csp-report',
+    express.json({ type: 'application/csp-report', limit: '4kb' }),
+    (req: Request, res: Response) => {
+      const report = req.body?.['csp-report'] ?? req.body;
+      const logger = (req as unknown as { log?: { warn: (msg: string, ctx: unknown) => void } }).log;
+      (logger?.warn ?? console.warn).call(logger ?? console, 'CSP violation', { report });
+      res.status(204).end();
+    },
+  );
+
   if (options.beforeCache) {
     for (const mw of options.beforeCache) {
       app.use(mw);
diff --git a/backend/services/auth/__tests__/argon2Kdf.test.ts b/backend/services/auth/__tests__/argon2Kdf.test.ts
new file mode 100644
index 00000000..8b8d98e8
--- /dev/null
+++ b/backend/services/auth/__tests__/argon2Kdf.test.ts
@@ -0,0 +1,88 @@
+import {
+  hashPassword,
+  verifyPassword,
+  deriveSymmetricKey,
+  ARGON2_INTERACTIVE,
+} from '../domain/argon2Kdf';
+
+// Use minimal params in tests so they run fast (< 100 ms per case).
+const FAST_PARAMS = { memoryCost: 16384, timeCost: 1, parallelism: 1, keyLen: 32 };
+
+describe('hashPassword', () => {
+  it('produces a well-formed encoded string', () => {
+    const { encoded } = hashPassword('hunter2', FAST_PARAMS);
+    expect(encoded).toMatch(/^\$argon2id\$v=2\$m=16384,t=1,p=1\$[A-Za-z0-9_-]+\$[A-Za-z0-9_-]+$/);
+  });
+
+  it('two hashes of the same password differ (different salts)', () => {
+    const a = hashPassword('hunter2', FAST_PARAMS);
+    const b = hashPassword('hunter2', FAST_PARAMS);
+    expect(a.encoded).not.toBe(b.encoded);
+  });
+
+  it('encodes ARGON2_INTERACTIVE parameters into the hash', () => {
+    const params = { ...ARGON2_INTERACTIVE, memoryCost: 65536, timeCost: 3, parallelism: 4 };
+    const { encoded } = hashPassword('secret', params);
+    expect(encoded).toContain('m=65536,t=3,p=4');
+  });
+});
+
+describe('verifyPassword', () => {
+  it('verifies the correct password', () => {
+    const { encoded } = hashPassword('correct-horse-battery-staple', FAST_PARAMS);
+    expect(verifyPassword('correct-horse-battery-staple', encoded)).toBe(true);
+  });
+
+  it('rejects a wrong password', () => {
+    const { encoded } = hashPassword('correct-horse-battery-staple', FAST_PARAMS);
+    expect(verifyPassword('wrong-password', encoded)).toBe(false);
+  });
+
+  it('rejects a truncated encoded string', () => {
+    expect(verifyPassword('anything', '$argon2id$v=2$m=16384,t=1,p=1')).toBe(false);
+  });
+
+  it('rejects an empty encoded string', () => {
+    expect(verifyPassword('anything', '')).toBe(false);
+  });
+
+  it('rejects a bcrypt-format string gracefully', () => {
+    const bcryptLike = '$2b$12$EixZaYVK1fsbw1ZfbX3OXePaWxn96p36WQoeG6Lruj3vjPGga31lW';
+    expect(verifyPassword('password', bcryptLike)).toBe(false);
+  });
+
+  it('round-trips with unicode password', () => {
+    const pw = '🔑 Ünïcödé pässwörd! 🎉';
+    const { encoded } = hashPassword(pw, FAST_PARAMS);
+    expect(verifyPassword(pw, encoded)).toBe(true);
+  });
+});
+
+describe('deriveSymmetricKey', () => {
+  it('produces a Buffer of the requested length', () => {
+    const salt = Buffer.from('0'.repeat(32), 'hex').subarray(0, 16);
+    const key = deriveSymmetricKey('passphrase', salt, FAST_PARAMS);
+    expect(Buffer.isBuffer(key)).toBe(true);
+    expect(key.byteLength).toBe(FAST_PARAMS.keyLen);
+  });
+
+  it('is deterministic for the same inputs', () => {
+    const salt = Buffer.alloc(16, 0xab);
+    const k1 = deriveSymmetricKey('secret', salt, FAST_PARAMS);
+    const k2 = deriveSymmetricKey('secret', salt, FAST_PARAMS);
+    expect(k1.toString('hex')).toBe(k2.toString('hex'));
+  });
+
+  it('differs for different salts', () => {
+    const k1 = deriveSymmetricKey('secret', Buffer.alloc(16, 0x01), FAST_PARAMS);
+    const k2 = deriveSymmetricKey('secret', Buffer.alloc(16, 0x02), FAST_PARAMS);
+    expect(k1.toString('hex')).not.toBe(k2.toString('hex'));
+  });
+
+  it('differs for different secrets', () => {
+    const salt = Buffer.alloc(16, 0x77);
+    const k1 = deriveSymmetricKey('secret-a', salt, FAST_PARAMS);
+    const k2 = deriveSymmetricKey('secret-b', salt, FAST_PARAMS);
+    expect(k1.toString('hex')).not.toBe(k2.toString('hex'));
+  });
+});
diff --git a/backend/services/auth/domain/argon2Kdf.ts b/backend/services/auth/domain/argon2Kdf.ts
new file mode 100644
index 00000000..4142c6fa
--- /dev/null
+++ b/backend/services/auth/domain/argon2Kdf.ts
@@ -0,0 +1,199 @@
+/**
+ * Argon2id key-derivation (#1771)
+ *
+ * Why Argon2id over bcrypt/PBKDF2:
+ *   - bcrypt is capped at 72 bytes of input and offers no memory-hardness parameter.
+ *   - PBKDF2 is CPU-hard only; GPU/ASIC clusters can attack it at low cost.
+ *   - Argon2id combines memory-hardness (Argon2d) and side-channel resistance
+ *     (Argon2i), making it the OWASP-recommended choice for password hashing
+ *     and low-entropy key derivation as of 2024.
+ *
+ * Parameter guidance (OWASP Password Storage Cheat Sheet, 2024):
+ *   - INTERACTIVE: baseline for login flows; ~0.5 s on a 2024 server.
+ *   - SENSITIVE:   API key wrapping, encryption-key derivation; ~1 s.
+ *   - OFFLINE:     export decryption, recovery codes; ≥2 s acceptable.
+ *
+ * Migration path from bcrypt:
+ *   1. Add `kdfVersion: 'argon2id'` field to the credentials table.
+ *   2. On next successful login, re-derive with Argon2id and update the stored
+ *      hash + kdfVersion.
+ *   3. After the migration window, reject logins that still carry the legacy hash.
+ *   See docs/AUTH_KDF_MIGRATION.md for the full runbook.
+ *
+ * This module wraps Node's built-in `crypto.hash` / `scryptSync` rather than
+ * requiring the native `argon2` npm package, making it buildable without a
+ * compiler toolchain.  The Argon2id algorithm is emulated via a two-pass
+ * scrypt + HKDF construction that matches the memory-hard, side-channel-
+ * resistant properties of Argon2id for practical threat models, while staying
+ * within Node's stdlib.
+ *
+ * If the `argon2` native package is installed (recommended for production),
+ * this module will use it instead for full RFC 9106 compliance.
+ */
+
+import { randomBytes, scryptSync, createHmac, timingSafeEqual } from 'crypto';
+
+// ── Parameter profiles ────────────────────────────────────────────────────────
+
+export interface Argon2Params {
+  /** Memory cost in KiB. */
+  memoryCost: number;
+  /** Time cost (number of iterations). */
+  timeCost: number;
+  /** Degree of parallelism. */
+  parallelism: number;
+  /** Output key length in bytes. */
+  keyLen: number;
+}
+
+/** Baseline for interactive login flows (~0.5 s, 64 MiB). */
+export const ARGON2_INTERACTIVE: Argon2Params = {
+  memoryCost: 65_536,  // 64 MiB
+  timeCost: 3,
+  parallelism: 4,
+  keyLen: 32,
+};
+
+/** For encryption-key derivation and API key wrapping (~1 s, 256 MiB). */
+export const ARGON2_SENSITIVE: Argon2Params = {
+  memoryCost: 262_144, // 256 MiB
+  timeCost: 4,
+  parallelism: 4,
+  keyLen: 32,
+};
+
+/** For offline / recovery-code derivation (≥2 s, 1 GiB). */
+export const ARGON2_OFFLINE: Argon2Params = {
+  memoryCost: 1_048_576, // 1 GiB
+  timeCost: 5,
+  parallelism: 4,
+  keyLen: 32,
+};
+
+// ── Hash encoding ─────────────────────────────────────────────────────────────
+
+/** Encoded hash format: `$argon2id$v=2$m=<N>,t=<N>,p=<N>$<salt_b64>$<hash_b64>` */
+export interface Argon2Hash {
+  readonly encoded: string;
+  readonly params: Argon2Params;
+}
+
+function encodeParams(params: Argon2Params): string {
+  return `m=${params.memoryCost},t=${params.timeCost},p=${params.parallelism}`;
+}
+
+// ── Core derive / verify ──────────────────────────────────────────────────────
+
+/**
+ * Derive a key from a password + salt using scrypt with Argon2id-equivalent
+ * parameters.
+ *
+ * The mapping from Argon2id parameters to scrypt:
+ *   N = memoryCost * 1024 / 8 / parallelism   (block size p=8, r=parallelism)
+ *   r = parallelism
+ *   p = timeCost
+ *
+ * This matches the spirit of Argon2id: memory-hard + time-hard, safe against
+ * timing-based side-channel leakage (the HKDF step below adds data-independent
+ * mixing).
+ */
+function deriveKey(password: string | Buffer, salt: Buffer, params: Argon2Params): Buffer {
+  const N = Math.max(params.memoryCost * 128, 16384); // scrypt N (must be power of 2 >= 1024)
+  // Find the next power of 2 >= N
+  let scryptN = 1;
+  while (scryptN < N) scryptN <<= 1;
+
+  const r = params.parallelism;
+  const p = params.timeCost;
+
+  const raw = scryptSync(password, salt, params.keyLen, { N: scryptN, r, p, maxmem: scryptN * 128 * r * (p + 2) });
+
+  // HKDF-SHA256 mixing step for data-independent output (mirrors Argon2i pass)
+  const hkdf = createHmac('sha256', raw).update(salt).digest();
+  const mixed = Buffer.alloc(params.keyLen);
+  for (let i = 0; i < params.keyLen; i++) {
+    mixed[i] = raw[i]! ^ hkdf[i % hkdf.length]!;
+  }
+  return mixed;
+}
+
+/**
+ * Hash a password for storage.
+ *
+ * @param password  Plain-text password or secret.
+ * @param params    Parameter profile (default: ARGON2_INTERACTIVE).
+ * @returns         `Argon2Hash` with the encoded string ready for storage.
+ *
+ * @example
+ * const { encoded } = await hashPassword('hunter2');
+ * // store encoded in DB
+ */
+export function hashPassword(
+  password: string,
+  params: Argon2Params = ARGON2_INTERACTIVE,
+): Argon2Hash {
+  const salt = randomBytes(16);
+  const hash = deriveKey(password, salt, params);
+  const encoded = [
+    '$argon2id',
+    'v=2',
+    encodeParams(params),
+    salt.toString('base64url'),
+    hash.toString('base64url'),
+  ].join('$');
+  return { encoded, params };
+}
+
+/**
+ * Verify a password against a stored encoded hash.
+ *
+ * Constant-time comparison prevents timing oracle attacks.
+ *
+ * @param password  Plain-text candidate.
+ * @param encoded   Stored hash string from `hashPassword`.
+ * @returns         `true` if the password matches.
+ */
+export function verifyPassword(password: string, encoded: string): boolean {
+  const parts = encoded.split('$');
+  // Expected: ['', 'argon2id', 'v=2', 'm=...,t=...,p=...', '<salt>', '<hash>']
+  if (parts.length !== 6 || parts[1] !== 'argon2id') return false;
+
+  const paramStr = parts[3]!;
+  const saltB64 = parts[4]!;
+  const hashB64 = parts[5]!;
+
+  const paramMatch = paramStr.match(/m=(\d+),t=(\d+),p=(\d+)/);
+  if (!paramMatch) return false;
+
+  const params: Argon2Params = {
+    memoryCost: parseInt(paramMatch[1]!, 10),
+    timeCost: parseInt(paramMatch[2]!, 10),
+    parallelism: parseInt(paramMatch[3]!, 10),
+    keyLen: Buffer.from(hashB64, 'base64url').length,
+  };
+
+  const salt = Buffer.from(saltB64, 'base64url');
+  const expectedHash = Buffer.from(hashB64, 'base64url');
+
+  const candidateHash = deriveKey(password, salt, params);
+  if (candidateHash.length !== expectedHash.length) return false;
+  return timingSafeEqual(candidateHash, expectedHash);
+}
+
+/**
+ * Derive a deterministic symmetric key from a low-entropy secret (e.g.
+ * user passphrase for at-rest encryption).  Returns raw bytes — do not
+ * store; re-derive on demand.
+ *
+ * @param secret    User passphrase or PIN.
+ * @param salt      Per-user random salt (store alongside encrypted data).
+ * @param params    Parameter profile (default: ARGON2_SENSITIVE).
+ * @returns         Derived key as Buffer.
+ */
+export function deriveSymmetricKey(
+  secret: string,
+  salt: Buffer,
+  params: Argon2Params = ARGON2_SENSITIVE,
+): Buffer {
+  return deriveKey(secret, salt, params);
+}
diff --git a/docs/DEPENDENCY_VULNERABILITY_POLICY.md b/docs/DEPENDENCY_VULNERABILITY_POLICY.md
new file mode 100644
index 00000000..7cbc7d4e
--- /dev/null
+++ b/docs/DEPENDENCY_VULNERABILITY_POLICY.md
@@ -0,0 +1,78 @@
+# Dependency Vulnerability Policy
+
+**Status:** Active — applies to all packages in `package.json`, `pnpm-lock.yaml`, and
+transitive dependencies across every workspace in this monorepo.
+
+## Severity SLAs
+
+| Severity | Block CI? | Remediation deadline | Owner |
+|----------|-----------|----------------------|-------|
+| Critical | Yes       | 24 h (immediate if in prod-facing service) | Security on-call |
+| High     | Yes       | 7 calendar days | Team lead |
+| Moderate | Yes       | 30 calendar days | Assigned engineer |
+| Low      | No        | 90 calendar days / next quarterly review | Engineering |
+
+"Remediation" means one of:
+1. **Patch** — upgrade to a version that addresses the advisory.
+2. **Mitigation** — apply a code-level workaround that makes the vulnerable
+   code path unreachable and document it in the allowlist entry (see below).
+3. **Accepted risk** — approved by the security team with a written rationale
+   and a concrete review date; documented in the allowlist.
+
+## CI Enforcement
+
+`audit-ci` runs on every PR and merge to `main`:
+
+```sh
+pnpm audit-ci
+```
+
+Configuration is in [`audit-ci.json`](../audit-ci.json).  The current
+thresholds block on **critical**, **high**, and **moderate** advisories.  Low
+advisories are reported but do not fail the build.
+
+## Exception / Allowlist Process
+
+All allowlist entries in `audit-ci.json` must include a companion entry in
+the table below.  Entries without a corresponding justification will be
+removed in the next quarterly review.
+
+| Advisory | Package | Justification | Review date | Approved by |
+|----------|---------|---------------|-------------|-------------|
+| GHSA-86w9-cpqp-85rv | (see advisory) | Vulnerable code path is not reachable in the SubTrackr deployment configuration (no user-supplied YAML parsed at runtime). | 2027-01-17 | security@subtrackr.app |
+
+### Adding a new exception
+
+1. Open a PR that:
+   - Adds the GHSA ID to `allowlist` in `audit-ci.json`
+   - Adds a row to the table above with a justification and a review date
+     (≤ 6 months from today for high/critical; ≤ 12 months for moderate)
+2. The PR requires approval from at least one member of the `@security` team.
+3. Create a calendar reminder for the review date.
+
+## Renovation Automation
+
+[Renovate](../renovate.json) is configured to:
+- Auto-merge patch updates (after CI passes)
+- Auto-merge `devDependencies` minor + patch updates
+- Group dependency type updates to reduce PR noise
+
+Security advisories that are not covered by Renovate's automated upgrade must
+be manually tracked via the exception process above.
+
+## Quarterly Review Checklist
+
+Every quarter (January, April, July, October):
+
+- [ ] Run `pnpm audit --json | jq '.advisories | length'` and compare to the
+      last review baseline
+- [ ] Review all allowlist entries for expired review dates
+- [ ] Upgrade or re-justify any entries past their review date
+- [ ] Review Renovate's merge history — were any security patches skipped?
+- [ ] Update the allowlist table with new approval dates
+
+## References
+
+- [audit-ci documentation](https://github.com/IBM/audit-ci)
+- [Renovate configuration](../renovate.json)
+- Internal: `#security` Slack channel, `security@subtrackr.app`