@oneuptime/common 12.0.21 → 12.0.23

package/Server/Utils/TwoFactorBackupCode.ts

@@ -0,0 +1,266 @@
+import { EncryptionSecret } from "../EnvironmentConfig";
+import ObjectID from "../../Types/ObjectID";
+import CaptureSpan from "./Telemetry/CaptureSpan";
+import crypto from "crypto";
+
+/*
+ * Single-use recovery codes for an account whose second factor is gone --
+ * the phone that held the authenticator app, or the security key that is now
+ * in a taxi somewhere.
+ *
+ * WHY THIS IS SERVER-ONLY
+ *
+ * Nothing in this file falls back to Math.random. Common/Utils/UUID does, on
+ * purpose, because it is bundled into the dashboard and has to keep producing
+ * well-formed ids in a browser with no Web Crypto -- which is exactly why
+ * ObjectID.generate() must never mint a code here. A backup code is a
+ * password-equivalent credential; if the platform cannot produce secure
+ * randomness, generation must throw rather than quietly emit something
+ * predictable. Common/Server/Utils/VerificationCode.ts makes the same call for
+ * the same reason.
+ *
+ * THE CODE SPACE
+ *
+ * Ten characters drawn uniformly from a 32 symbol alphabet is 2^50 codes, and
+ * a user holds ten of them at once -- so a blind guess lands with probability
+ * ~10/2^50, about one in 10^14. That is far beyond anything the rate limiter
+ * needs to defend, which matters because it settles the design question a TOTP
+ * code cannot settle: the six digit space is small enough that the limiter IS
+ * the control, whereas here the code itself is.
+ *
+ * THE HASH
+ *
+ * Codes are stored as HMAC-SHA256 keyed by the instance's EncryptionSecret,
+ * never in the clear, and the fast keyed-digest lane is the RIGHT one here
+ * rather than scrypt. scrypt exists to make guessing a low-entropy,
+ * human-chosen secret expensive; a code minted above has no low-entropy
+ * structure to guess, so the cost would buy nothing and would be paid on every
+ * verification. Keying with the EncryptionSecret is what a bare SHA-256 would
+ * miss: it lives in configuration rather than in Postgres, so a database dump
+ * on its own cannot be run through a dictionary of every possible code.
+ *
+ * The digest is deterministic given (userId, code), which is deliberate and is
+ * what makes single-use consumption a single conditional UPDATE rather than a
+ * read of every one of the user's rows followed by a comparison of each. A
+ * per-row salt would forfeit that for no gain -- see the note on domain
+ * separation below for the property it would have been bought for, which the
+ * userId already provides.
+ */
+
+/**
+ * Crockford's Base32 alphabet: the digits and the uppercase letters, minus
+ * I, L, O and U.
+ *
+ * These codes get read off a screen and typed back in months later, possibly
+ * from a piece of paper in a drawer, so the alphabet is chosen for the eye
+ * rather than for density. I/1, L/1 and O/0 are the pairs people transcribe
+ * wrongly; U is dropped by Crockford so that no code can spell an obscenity
+ * at a user who did nothing to deserve one.
+ *
+ * Exactly 32 symbols, so each character carries a clean five bits and the
+ * entropy arithmetic in the header comment is exact rather than approximate.
+ */
+export const BackupCodeAlphabet: string = "0123456789ABCDEFGHJKMNPQRSTVWXYZ";
+
+/** Characters per code. Ten symbols over a 32 symbol alphabet is 2^50. */
+export const BackupCodeLength: number = 10;
+
+/**
+ * How many codes a user is issued at once.
+ *
+ * Ten is enough that losing a phone does not become an emergency after the
+ * second sign-in, and small enough that the printed list is one short block a
+ * person will actually keep.
+ */
+export const BackupCodeSetSize: number = 10;
+
+/** Characters per group in the displayed form, e.g. `AB3D5-9XZQ2`. */
+const DISPLAY_GROUP_LENGTH: number = 5;
+
+/*
+ * Prefixed into every digest so a stored hash is bound to the scheme that
+ * produced it. If the construction below ever has to change, old rows keep
+ * verifying under the version they were written with instead of silently
+ * failing to match and locking a user out of their own recovery codes.
+ */
+const HASH_SCHEME_VERSION: string = "v1";
+
+/*
+ * Characters a person plausibly types in place of a symbol that is not in the
+ * alphabet. Applied before the strip below, so `O` becomes `0` rather than
+ * being deleted -- deleting it would shorten the code and guarantee a
+ * mismatch, which is the confusing failure this map exists to avoid.
+ */
+const AMBIGUOUS_CHARACTER_MAP: Record<string, string> = {
+  I: "1",
+  L: "1",
+  O: "0",
+};
+
+export default class TwoFactorBackupCode {
+  /**
+   * One code, drawn uniformly from the alphabet above.
+   *
+   * `crypto.randomInt` per character rather than `randomBytes(n) % 32`. The
+   * modulo version happens to be uniform for this alphabet only because 32
+   * divides 256, and it stops being uniform the moment somebody edits the
+   * alphabet -- silently, with no test that would notice. `randomInt` rejects
+   * out-of-range draws internally for whatever bound it is given, so the
+   * uniformity does not depend on a coincidence nobody wrote down.
+   */
+  @CaptureSpan()
+  public static generateCode(): string {
+    let code: string = "";
+
+    for (let index: number = 0; index < BackupCodeLength; index++) {
+      code += BackupCodeAlphabet.charAt(
+        crypto.randomInt(0, BackupCodeAlphabet.length),
+      );
+    }
+
+    return code;
+  }
+
+  /**
+   * A full set of distinct codes.
+   *
+   * Duplicates are not a security problem -- at 2^50 the birthday odds across
+   * ten draws are around 4e-14 -- but a duplicate WOULD be a correctness
+   * problem downstream: two rows sharing a digest means consuming one leaves a
+   * second, identical, still-valid code behind, so a "single-use" code would
+   * work twice. Cheaper to rule out here than to reason about there.
+   */
+  @CaptureSpan()
+  public static generateCodeSet(
+    count: number = BackupCodeSetSize,
+  ): Array<string> {
+    const codes: Set<string> = new Set<string>();
+
+    while (codes.size < count) {
+      codes.add(TwoFactorBackupCode.generateCode());
+    }
+
+    return Array.from(codes);
+  }
+
+  /**
+   * The form shown to the user: one hyphen in the middle, for the same reason
+   * the alphabet drops ambiguous letters -- a ten character run is hard to
+   * read back without losing your place.
+   *
+   * Purely cosmetic. `normalizeCode` strips the hyphen straight back out, so
+   * a user may type the code with it, without it, or with the spaces their
+   * password manager pasted in.
+   */
+  @CaptureSpan()
+  public static formatForDisplay(code: string): string {
+    const groups: Array<string> = [];
+
+    for (
+      let index: number = 0;
+      index < code.length;
+      index += DISPLAY_GROUP_LENGTH
+    ) {
+      groups.push(code.substring(index, index + DISPLAY_GROUP_LENGTH));
+    }
+
+    return groups.join("-");
+  }
+
+  /**
+   * Reduce whatever the user typed to the canonical form the digest is
+   * computed over.
+   *
+   * Everything here is about not rejecting somebody who supplied exactly the
+   * right secret material: the display hyphen, the spaces a clipboard adds,
+   * lowercase from a phone keyboard, and the three transcription confusions
+   * the alphabet was chosen to make survivable. Anything still outside the
+   * alphabet after that is dropped rather than rejected -- this is a
+   * canonicaliser, not a validator; the digest comparison is what decides
+   * whether a code is real.
+   *
+   * @param rawCode - The code exactly as submitted.
+   * @returns The code reduced to alphabet symbols, uppercase.
+   */
+  @CaptureSpan()
+  public static normalizeCode(rawCode: string): string {
+    /*
+     * The code arrives straight off a JSON body, so it is only a string
+     * because the client chose to send one. A number or an object here must
+     * fail verification, not throw out of `.toUpperCase` and surface as a 500.
+     */
+    if (typeof rawCode !== "string" || !rawCode) {
+      return "";
+    }
+
+    let normalized: string = "";
+
+    for (const character of rawCode.toUpperCase()) {
+      const mapped: string = AMBIGUOUS_CHARACTER_MAP[character] || character;
+
+      if (BackupCodeAlphabet.includes(mapped)) {
+        normalized += mapped;
+      }
+    }
+
+    return normalized;
+  }
+
+  /**
+   * The digest stored for one of `userId`'s codes.
+   *
+   * Domain separated by the owning user, which is what stops one precomputed
+   * table from inverting every account's codes at once and stops two users who
+   * happen to be issued the same code from being visibly linked by a matching
+   * row. The user id is known at verification time -- the password has already
+   * been accepted by then -- so binding to it costs nothing.
+   *
+   * Both parts are LENGTH-PREFIXED into the message rather than merely
+   * concatenated, so no pair of (userId, code) values can be rearranged into
+   * the same byte string as another pair. Concatenation alone is a real
+   * ambiguity, not a theoretical one: without prefixes, ("ab", "cd") and
+   * ("abc", "d") hash identically.
+   */
+  @CaptureSpan()
+  public static hashCode(data: { code: string; userId: ObjectID }): string {
+    const message: string = [
+      HASH_SCHEME_VERSION,
+      data.userId.toString(),
+      TwoFactorBackupCode.normalizeCode(data.code),
+    ]
+      .map((part: string) => {
+        return `${part.length}:${part}`;
+      })
+      .join("");
+
+    return crypto
+      .createHmac("sha256", EncryptionSecret.toString())
+      .update(message)
+      .digest("hex");
+  }
+
+  /**
+   * Compare two digests without leaking, through timing, how many leading
+   * characters matched.
+   *
+   * The consume path compares digests in Postgres rather than here, so this is
+   * for callers that already hold both -- but it exists so that nobody writing
+   * one reaches for `===` and hands an attacker a way to recover a stored hash
+   * one character at a time.
+   *
+   * Lengths are compared first because `crypto.timingSafeEqual` THROWS on a
+   * length mismatch. Digests here are always 64 hex characters, so an early
+   * exit only happens on malformed input, where the length is not the secret.
+   */
+  @CaptureSpan()
+  public static isHashEqual(a: string, b: string): boolean {
+    if (!a || !b || a.length !== b.length) {
+      return false;
+    }
+
+    return crypto.timingSafeEqual(
+      Buffer.from(a, "utf8"),
+      Buffer.from(b, "utf8"),
+    );
+  }
+}

package/Tests/Server/API/UserAuthenticationAPI.test.ts

@@ -96,6 +96,14 @@ const AUTHENTICATION_STATUS: UserAuthenticationStatus = {
    */
   twoFactorAuthStatus: TwoFactorAuthStatus.EnabledPendingSetup,
   verifiedTwoFactorAuthMethodCount: 0,
+
+  /*
+   * Reported alongside the method count, never folded into it -- backup codes
+   * are the way back in to an account whose factor is unreachable, not a
+   * factor. Zero here keeps this fixture the pure "mandated, nothing set up"
+   * case it exists to be.
+   */
+  unusedTwoFactorBackupCodeCount: 0,
   hasPendingPasswordResetLink: false,
 };
 
@@ -201,9 +209,13 @@ describe("GET /user/:userId/authentication-status", () => {
 
   test("sends exactly what the service reported", async () => {
     /*
-     * The page renders these four booleans directly. A field dropped or
-     * renamed on the way out reads as `undefined`, which the UI shows as "no
-     * password" for an account that has one.
+     * The page renders these fields directly. One dropped or renamed on the
+     * way out reads as `undefined`, which the UI shows as "no password" for an
+     * account that has one.
+     *
+     * Asserted as an EXACT object rather than field by field, so a new field
+     * added to the DTO has to be added here too -- which is how this test has
+     * done its job twice now, most recently for the backup code count.
      */
     await callRoute({
       method: "GET",
@@ -217,6 +229,7 @@ describe("GET /user/:userId/authentication-status", () => {
       isTwoFactorAuthEnabled: true,
       twoFactorAuthStatus: TwoFactorAuthStatus.EnabledPendingSetup,
       verifiedTwoFactorAuthMethodCount: 0,
+      unusedTwoFactorBackupCodeCount: 0,
       hasPendingPasswordResetLink: false,
     });
   });

package/Tests/Server/API/UserTwoFactorAuthAdminAPI.test.ts

@@ -792,6 +792,7 @@ describe("GET /user/:userId/authentication-status two factor auth fields", () =>
     isTwoFactorAuthEnabled: true,
     twoFactorAuthStatus: TwoFactorAuthStatus.EnabledPendingSetup,
     verifiedTwoFactorAuthMethodCount: 0,
+    unusedTwoFactorBackupCodeCount: 0,
     hasPendingPasswordResetLink: false,
   };
 
@@ -826,6 +827,7 @@ describe("GET /user/:userId/authentication-status two factor auth fields", () =>
       isTwoFactorAuthEnabled: true,
       twoFactorAuthStatus: TwoFactorAuthStatus.EnabledPendingSetup,
       verifiedTwoFactorAuthMethodCount: 0,
+      unusedTwoFactorBackupCodeCount: 0,
       hasPendingPasswordResetLink: false,
     });
   });
@@ -863,6 +865,9 @@ describe("GET /user/:userId/authentication-status two factor auth fields", () =>
 
     expect(Object.keys(payload)).toContain("verifiedTwoFactorAuthMethodCount");
     expect(payload["verifiedTwoFactorAuthMethodCount"]).toBe(0);
+
+    expect(Object.keys(payload)).toContain("unusedTwoFactorBackupCodeCount");
+    expect(payload["unusedTwoFactorBackupCodeCount"]).toBe(0);
   });
 
   test("reports a configured account with its real method count", async () => {
@@ -871,14 +876,17 @@ describe("GET /user/:userId/authentication-status two factor auth fields", () =>
      * operator would reset a user who was perfectly fine and lock them out
      * themselves.
      */
-    statusSpy.mockResolvedValue({
+    const configuredStatus: UserAuthenticationStatus = {
       hasPassword: false,
       isEmailVerified: true,
       isTwoFactorAuthEnabled: true,
       twoFactorAuthStatus: TwoFactorAuthStatus.EnabledConfigured,
       verifiedTwoFactorAuthMethodCount: 2,
+      unusedTwoFactorBackupCodeCount: 5,
       hasPendingPasswordResetLink: false,
-    } as UserAuthenticationStatus);
+    };
+
+    statusSpy.mockResolvedValue(configuredStatus);
 
     await callRoute({
       method: "GET",
@@ -892,6 +900,14 @@ describe("GET /user/:userId/authentication-status two factor auth fields", () =>
       TwoFactorAuthStatus.EnabledConfigured,
     );
     expect(payload["verifiedTwoFactorAuthMethodCount"]).toBe(2);
+
+    /*
+     * Carried alongside the method count, never folded into it. This is the
+     * number that tells an operator whether the user can get back in without
+     * them -- so a reset that would sign the user out everywhere is only the
+     * right button when it is zero.
+     */
+    expect(payload["unusedTwoFactorBackupCodeCount"]).toBe(5);
   });
 
   test("sets the no-cache headers before the payload goes out", async () => {