| パッケージ | 影響を受けるバージョン | 修正済みバージョン |
|---|---|---|
| github.com/ubuntu/authd | ≤ 0.0.0-20230706090440-d8cb2d561419< 0.3.6 | 0.3.6 |
| authd | < 0.3.6 | 0.3.6 |
CVE description:
Authd, through version 0.3.6, did not sufficiently randomize user IDs to prevent collisions. A local attacker who can register user names could spoof another user's ID and gain their privileges.
----- original report -----
authd assigns user IDs as a pure function of the user name. Moreover, the set of UIDs is much too small for pseudo-random assignment to work: the birthday bound predicts random collisions will occur with probability 50% after only 54 562 IDs were assigned.
authd only checks for uniqueness within its local cache, which
/var/cache ;The current GenerateID method, authored in September 2024 (commit a6c85ed24b8d17a2d11c859e8d70f5a52fa69690),
repeatedly hashes the user name until the 4 leading bytes fall into the interval [60 000; 2³¹[ :
https://github.com/ubuntu/authd/blob/f9f851540e6377fca18a45ce7a02d024c1dbd6e9/internal/users/manager.go#L425
https://github.com/ubuntu/authd/blob/f9f851540e6377fca18a45ce7a02d024c1dbd6e9/internal/services/nss/nss.go#L188
Previous versions are affected by similar issues, though without the use of a cryptographic hash in GenerateID, making exploitation computationally-easier.
Since GenerateID is a pure function with no secret input, and the set of UIDs is small, an adversary which can register users with chosen names can
authd, or a system user whose well-known ID is in a range which overlaps authd's.In the latter case, as all access control performed by the Linux kernel (and other Unices' kernels) is based on IDs and not usernames, if the attacker can sign into a system, they will have the same privileges as the target user. The attacker can bypass the uniqueness check in (at least) the following ways:
これらの製品をお使いですか？サイトをスキャンして、脆弱なバージョンが公開されていないか確認しましょう。
無料でサイトをスキャン/var/cache ;authd's range ;alice is known to log into 1.example.com ;bob ;bob and logs into 2.example.com, succeeding if alice hasn't (recently) logged into that system ;2 as if they were alice; assuming /home is shared, they could manipulate ~alice/.ssh/authorized_keys, ~alice/.config, alice's shell's initialization file, etc.
Note: NFSv4's idmap mechanism may prevent this, but isn't enabled by default (unless Kerberos is used, which isn't the case in an authd deployment)1.example.com is usually trivial.Since the necessary computation can be performed entirely offline, this wouldn't be affected by any rate-limits, and the only audit trail would be a single user registration. This would require on average less than 2³¹ computations of GenerateID: assuming SHA-256's cost is 25 cycles-per-byte, a clock speed of 3GHz, and short (≤32B) generated usernames, this is less than 10 minutes of a single core's time.
The simplest and likely-best remediation path would be for the external IdP to provide a guaranteed-unique user ID in the correct range. In OIDC, this is commonly communicated through a claim, though its name would need to be configurable as there's no real standard:
cern_person_id: https://auth.docs.cern.ch/user-documentation/oidc/config/ ;uid or uidNumber ;This is also supported by other commonplace identity providers, such as LDAP and Active Directory: https://learn.microsoft.com/en-us/windows/win32/adschema/a-uidNumber
MS Entra presumably supports this as well.
If that is not possible for some reason, architectural changes to authd would likely be required: assigning user IDs from a small space (such as Linux's 32b UIDs) requires mutable state to ensure uniqueness, whereas authd's design currently assumes no mutable state is held, aside from some transient, local cache. Moreover, that mutable state may need to be synchronised across multiple machines as uniform UIDs are often necessary, for instance when accessing a common networked filesystem.
Thanks to Michael Gebetsroither for assisting with the writeup, and Jamie Bliss for the same as well as investigating when the issue was introduced in authd.