GHSA-ghmh-jhmj-wcmf

internal/store/sqlite.go:1177,1192,1221,1245 — the enrollment_tokens.token column holds the raw UUID token. ConsumeToken does WHERE token = ? against the raw string. Compare with operator_api_keys.key_hash, which is SHA-256 hex (constructed in internal/api/middleware.go:51-53).

Affected

All released versions up to v0.3.0.

Threat model

Read access to nebula-mgmt.db: backup, snapshot, file-system access, future SQL-injection sink. The principle of defense-in-depth: API keys are hashed at rest; enrollment tokens — which grant the same lifecycle authority over a host's identity — are not.

An attacker who reads the DB before a legitimate agent enrolls can consume the single-use token first, mint a cert against their own keypair, and take the agent's intended Nebula identity.

Suggested fix

1. Schema migration: rename enrollment_tokens.token → token_hash (or add the new column and drop the old after backfill of pending rows).
2. Store SHA-256 of token on create:

   sum := sha256.Sum256([]byte(token))
   row.TokenHash = hex.EncodeToString(sum[:])
   

3. ConsumeToken accepts the raw token, hashes once, looks up by hash, atomically marks consumed.

Side bonus: take this opportunity to switch the token format from uuid.New().String() (122 bits) to hex.EncodeToString(crypto/rand 32 bytes) (256 bits), matching the project's session-token and API-key conventions. UUIDs are recognisable in logs and crash dumps; opaque hex blends in.

TOTP recovery codes appear to already be SHA-256 hashed at rest (internal/web/totp.go:74-78) — confirming that pattern is intentional elsewhere, just missed here.

internal/store/sqlite.go:1177,1192,1221,1245 — the enrollment_tokens.token column holds the raw UUID token. ConsumeToken does WHERE token = ? against the raw string. Compare with operator_api_keys.key_hash, which is SHA-256 hex (constructed in internal/api/middleware.go:51-53).

Affected

All released versions up to v0.3.0.

Threat model

Read access to nebula-mgmt.db: backup, snapshot, file-system access, future SQL-injection sink. The principle of defense-in-depth: API keys are hashed at rest; enrollment tokens — which grant the same lifecycle authority over a host's identity — are not.

An attacker who reads the DB before a legitimate agent enrolls can consume the single-use token first, mint a cert against their own keypair, and take the agent's intended Nebula identity.

Suggested fix

1. Schema migration: rename enrollment_tokens.token → token_hash (or add the new column and drop the old after backfill of pending rows).
2. Store SHA-256 of token on create:

   sum := sha256.Sum256([]byte(token))
   row.TokenHash = hex.EncodeToString(sum[:])
   

3. ConsumeToken accepts the raw token, hashes once, looks up by hash, atomically marks consumed.

Side bonus: take this opportunity to switch the token format from uuid.New().String() (122 bits) to hex.EncodeToString(crypto/rand 32 bytes) (256 bits), matching the project's session-token and API-key conventions. UUIDs are recognisable in logs and crash dumps; opaque hex blends in.

TOTP recovery codes appear to already be SHA-256 hashed at rest (internal/web/totp.go:74-78) — confirming that pattern is intentional elsewhere, just missed here.

nebula-mesh's stores enrollment tokens unhashed in SQLite in github.com/juev/nebula-mesh