| name | entra-token-verification |
| description | Verify a Microsoft Entra ID (Azure AD) access token server-side and derive identity from its claims. Use when adding or reviewing Entra/Azure AD JWT verification, when a verifier accepts tokens it should reject (or rejects tokens it should accept), or when binding a verified token to an identity record (GUID/oid comparison, netid derivation). Covers the v2-issuer/audience trap, the ID-token-as-access-token trap, the missing-email-claim trap, and case-folding hazards in identity comparison. |
Entra Token Verification
Tier: spec anchored to live code (second consumer adapted the original — the
invariants below are the value, not a shared package; regenerate the implementation
per consumer, per CONTRIBUTING.md).
Verifying an Entra access token server-side (via jose + the tenant JWKS endpoint) is
easy to get subtly, silently wrong: a misconfigured issuer/audience check passes v1
tokens that should be rejected, or a verifier that never checks for scp/roles
accepts an ID token as if it were an access token. These are the constraints that keep
recurring across consumers — freeze them, regenerate the verification code itself.
The constraints
-
accessTokenAcceptedVersion must be 2. The v2-issuer check
(https://login.microsoftonline.com/<tenant>/v2.0) rejects every token if the app
registration still issues v1 tokens — v1 tokens carry
iss https://sts.windows.net/<tenant>/, which never matches.
verify: az ad app show --id <appId> --query api.requestedAccessTokenVersion
expect: 2
checked: 2026-08-01
-
Audience must accept BOTH the bare clientId and api://<clientId>. v2 tokens
carry either form depending on how the scope was requested (api://<clientId>/.default
vs. <clientId>/.default); accepting only one silently rejects the other.
verify: decode a real token's aud claim (base64url-decode the JWT's middle segment)
and confirm which form the tenant actually issues for this app.
checked: 2026-08-01 (both quick and mykai-portal issue the bare-clientId form; accept
both anyway — do not narrow to what's currently observed).
-
Reject tokens carrying neither scp nor roles. An access-token verifier that
only checks signature/issuer/audience will also accept a valid ID token presented
as an access token — ID tokens pass every one of those checks but were never scoped
for API access. Real access tokens carry scp (delegated permissions) or roles
(app permissions); ID tokens carry neither.
verify: acquire an ID token deliberately (e.g. via the auth-code flow's id_token
response field) and confirm the verifier throws on it.
checked: 2026-08-01
-
The email claim is frequently ABSENT from Entra v2 access tokens. Identity must
be derived from preferred_username (the UPN), not email. This is a live failure
mode, not a theoretical one.
verify: decode a real access token's payload and check for the email key.
checked: 2026-06 (mindforum production failure — verifier assumed email present,
broke for real users whose tokens omitted it).
-
GUID/oid comparison must be case-insensitive. GUIDs are case-insensitive per
RFC 4122; a case-sensitive compare against an append-only identity ledger (e.g. a
provisioning registry's history) permanently locks out a legitimate user the moment
Entra or any client normalizes casing differently than the stored record.
verify: compare a stored oid and a live token's oid with .toLowerCase() on both
sides before ===; confirm a differently-cased-but-equal pair matches.
checked: 2026-08-01 (mykai portal review cycle, identity.mjs)
-
Case-fold AFTER rejecting non-ASCII, never before. U+212A KELVIN SIGN lowercases
to ASCII k under .toLowerCase(), so a crafted non-ASCII UPN/netid can collide with
a distinct ASCII identifier once folded. Reject any non-printable-ASCII input
(^[\x20-\x7E]+$) before calling .toLowerCase() on anything used for identity
comparison — this removes the whole confusable-codepoint class rather than trying to
enumerate it.
verify: feed K...@illinois.edu through the netid/identity comparison path and
confirm it is rejected as non-ASCII, not folded and matched.
checked: 2026-08-01 (mykai portal review cycle, identity.mjs)
Failure modes if skipped
- Skip (1) or (2): the verifier throws on legitimate tokens (loud, but wastes a debugging
session per consumer since the error is a generic issuer/audience mismatch).
- Skip (3): silent privilege escalation — an ID token grants access it was never scoped for.
- Skip (4): silent identity-resolution failure for a subset of real users (mindforum: this
shipped to production before being caught).
- Skip (5) or (6): silent identity confusion — either locks out a legitimate user forever
(case-sensitive GUID compare against an append-only ledger) or lets one identifier
impersonate another (Kelvin-sign collision).
Anchors
~/code/quick/src/auth/entra.ts — original, TypeScript, jose-based verification
(verifyEntraToken, mapClaims). Constraints 1-4 live here as inline comments at the
point they matter.
~/code/mykai/portal/lib/entra.mjs — second consumer, direct JS port
of the above (same constraints 1-4, unchanged).
~/code/mykai/portal/lib/identity.mjs — second consumer's hardening
beyond the port: constraints 5-6 (bindIdentity, netidFromUpn), plus fail-closed
handling of malformed/missing history and epoch-scoped oid binding (a provisioning-
registry-specific invariant, not a general Entra constraint — read the file's own
comments if adapting a similar ledger).
Checked: 2026-08-01.
Related