CVE-2025-0928

In Juju versions prior to 3.6.8 and 2.9.52, any authenticated controller user was allowed to upload arbitrary agent binaries to any model or to the controller itself, without verifying model membership or requiring explicit permissions. This enabled the distribution of poisoned binaries to new or upgraded machines, potentially resulting in remote code execution.

In Juju versions prior to 3.6.8 and 2.9.52, any authenticated controller user was allowed to upload arbitrary agent binaries to any model or to the controller itself, without verifying model membership or requiring explicit permissions. This enabled the distribution of poisoned binaries to new or upgraded machines, potentially resulting in remote code execution.

Summary

You can affect the agent binaries used in a Juju controller and the code that is run in the binaries by simply having a user account on a controller. You aren't required to have a model or any permissions. This just requires a user account in the controller database.

Details

Because of the way Juju upload tools code works in the controller it only checks that the user uploading agent binaries is authenticated and is a user tag. No more checks are performed and it allows that user to upload binaries to any model they like (as long as they know the model uuid) or upload binaries to the controller (attacker doesn't need to know any uuid's for controller or controller model).

Once the poison binaries have been uploaded any new machine that is started in the affected model or controller will get started with the poison binaries. Alternatively administrator's of the controller running either juju upgrade-controller or juju upgrade-model will force distribution of the poisoned binaries to all machines in either the model or poison the controllers themselves.

On top of this the exploit can be done with the Juju client tooling itself and no real knowledge on constructing raw API requests is required.

The tools handler is the main piece of code that is used in the APIServer for handling upload requests and persisting the data uploaded: The following code references is how Juju uses and defines this:

• The tools upload handler is defined here (https://github.com/juju/juju/blob/3.6/apiserver/apiserver.go#L972)
• The tools upload handler is created in the api server here (https://github.com/juju/juju/blob/4bcbd094097016b2fde926afd8c9e590eabb3f0c/apiserver/apiserver.go#L766C2-L766C25).
• The main authoriser that is used for the upload handler is created here (https://github.com/juju/juju/blob/4bcbd094097016b2fde926afd8c9e590eabb3f0c/apiserver/apiserver.go#L770C2-L770C28)
• The upload handler is registered for the model here (https://github.com/juju/juju/blob/4bcbd094097016b2fde926afd8c9e590eabb3f0c/apiserver/apiserver.go#L902)
• The upload handler is registered for the controller here (https://github.com/juju/juju/blob/4bcbd094097016b2fde926afd8c9e590eabb3f0c/apiserver/apiserver.go#L972)

The authoriser that is used (https://github.com/juju/juju/blame/4bcbd094097016b2fde926afd8c9e590eabb3f0c/apiserver/httpcontext.go#L209) only confirms that the logged in user is authenticated and authenticated as a user tag. No other checks are performed.

The toolsUploaderHandler also uses another server func for getting the Mongo state. This also confirms a logged in user but the state that is returned to the caller is scoped to whatever model the requester has asked for. No checks are performed to make sure that the user in question actually has access to this model or the controller. See code here (https://github.com/juju/juju/blob/4e50a28cdde17832aa31634915fbe7442dca6ab3/apiserver/httpcontext.go#L38). We end up here through a few layers of indirection of https://github.com/juju/juju/blob/4bcbd094097016b2fde926afd8c9e590eabb3f0c/apiserver/apiserver.go#L768

We can also see that when handlers are registered with no model uuid scope in the handler like the controller registration of the tools upload handler, the model uuid gets defaulted to that of the controller model. See (https://github.com/juju/juju/blob/4bcbd094097016b2fde926afd8c9e590eabb3f0c/apiserver/apiserver.go#L690).

PoC

This proof of concept was done with the latest tip of the juju/juju 3.6 branch (https://github.com/juju/juju/commit/cd12b4951d657a980e113564bf2ea82f167589fd). Pull this code and work from inside of the root of the code base. It is expected that this security issue applies to 2.9 onwards as well.

Repo steps:

1. Bootstrap a new controller to lxd. This was done with a compiled client from the branch but there is no reason performing this action from latest snap won't produce the same result. juju bootstrap localhost sec-demo

2. Add a new…

In Juju versions prior to 3.6.8 and 2.9.52, any authenticated controller user was allowed to upload arbitrary agent binaries to any model or to the controller itself, without verifying model membership or requiring explicit permissions. This enabled the distribution of poisoned binaries to new or upgraded machines, potentially resulting in remote code execution.

In Juju versions prior to 3.6.8 and 2.9.52, any authenticated controller user was allowed to upload arbitrary agent binaries to any model or to the controller itself, without verifying model membership or requiring explicit permissions. This enabled the distribution of poisoned binaries to new or upgraded machines, potentially resulting in remote code execution.

Summary

You can affect the agent binaries used in a Juju controller and the code that is run in the binaries by simply having a user account on a controller. You aren't required to have a model or any permissions. This just requires a user account in the controller database.

Details

Because of the way Juju upload tools code works in the controller it only checks that the user uploading agent binaries is authenticated and is a user tag. No more checks are performed and it allows that user to upload binaries to any model they like (as long as they know the model uuid) or upload binaries to the controller (attacker doesn't need to know any uuid's for controller or controller model).

Once the poison binaries have been uploaded any new machine that is started in the affected model or controller will get started with the poison binaries. Alternatively administrator's of the controller running either juju upgrade-controller or juju upgrade-model will force distribution of the poisoned binaries to all machines in either the model or poison the controllers themselves.

On top of this the exploit can be done with the Juju client tooling itself and no real knowledge on constructing raw API requests is required.

The tools handler is the main piece of code that is used in the APIServer for handling upload requests and persisting the data uploaded: The following code references is how Juju uses and defines this:

• The tools upload handler is defined here (https://github.com/juju/juju/blob/3.6/apiserver/apiserver.go#L972)
• The tools upload handler is created in the api server here (https://github.com/juju/juju/blob/4bcbd094097016b2fde926afd8c9e590eabb3f0c/apiserver/apiserver.go#L766C2-L766C25).
• The main authoriser that is used for the upload handler is created here (https://github.com/juju/juju/blob/4bcbd094097016b2fde926afd8c9e590eabb3f0c/apiserver/apiserver.go#L770C2-L770C28)
• The upload handler is registered for the model here (https://github.com/juju/juju/blob/4bcbd094097016b2fde926afd8c9e590eabb3f0c/apiserver/apiserver.go#L902)
• The upload handler is registered for the controller here (https://github.com/juju/juju/blob/4bcbd094097016b2fde926afd8c9e590eabb3f0c/apiserver/apiserver.go#L972)

The authoriser that is used (https://github.com/juju/juju/blame/4bcbd094097016b2fde926afd8c9e590eabb3f0c/apiserver/httpcontext.go#L209) only confirms that the logged in user is authenticated and authenticated as a user tag. No other checks are performed.

The toolsUploaderHandler also uses another server func for getting the Mongo state. This also confirms a logged in user but the state that is returned to the caller is scoped to whatever model the requester has asked for. No checks are performed to make sure that the user in question actually has access to this model or the controller. See code here (https://github.com/juju/juju/blob/4e50a28cdde17832aa31634915fbe7442dca6ab3/apiserver/httpcontext.go#L38). We end up here through a few layers of indirection of https://github.com/juju/juju/blob/4bcbd094097016b2fde926afd8c9e590eabb3f0c/apiserver/apiserver.go#L768

We can also see that when handlers are registered with no model uuid scope in the handler like the controller registration of the tools upload handler, the model uuid gets defaulted to that of the controller model. See (https://github.com/juju/juju/blob/4bcbd094097016b2fde926afd8c9e590eabb3f0c/apiserver/apiserver.go#L690).

PoC

This proof of concept was done with the latest tip of the juju/juju 3.6 branch (https://github.com/juju/juju/commit/cd12b4951d657a980e113564bf2ea82f167589fd). Pull this code and work from inside of the root of the code base. It is expected that this security issue applies to 2.9 onwards as well.

Repo steps:

1. Bootstrap a new controller to lxd. This was done with a compiled client from the branch but there is no reason performing this action from latest snap won't produce the same result. juju bootstrap localhost sec-demo

2. Add a new…

Juju allows arbitrary executable uploads via authenticated endpoint without authorization in github.com/juju/juju.

NOTE: The source advisory for this report contains additional versions that could not be automatically mapped to standard Go module versions.

(If this is causing false-positive reports from vulnerability scanners, please suggest an edit to the report.)

The additional affected modules and versions are: github.com/juju/juju before v0.0.0-20250619215741-4034aa13c7cf.

These are all security issues fixed in the govulncheck-vulndb-0.0.20250730T213748-1.1 package on the GA media of openSUSE Tumbleweed.