(adapted from information on Template plugin in CloudBees Plugins guide)
Various Jenkins plugins require that users define custom scripts, most commonly in the Groovy language, to customize Jenkins’s behavior. If everyone who writes these scripts is a Jenkins administrator—specifically if they have the Overall/RunScripts permission, used for example by the Script Console link—then they can write whatever scripts they like. These scripts may directly refer to internal Jenkins objects using the same API offered to plugins. Such users must be completely trusted, as they can do anything to Jenkins (even changing its security settings or running shell commands on the server).
However, if some script authors are “regular users” with only more limited permissions, such as Job/Configure, it is inappropriate to let them run arbitrary scripts. To support such a division of roles, the Script Security library plugin can be integrated into various feature plugins. It supports two related systems: script approval, and Groovy sandboxing.
The first, and simpler, security system is to allow any kind of script to be run, but only with an administrator’s approval. There is a globally maintained list of approved scripts which are judged to not perform any malicious actions.
When an administrator saves some kind of configuration (for example, a job), any scripts it contains are automatically added to the approved list. They are ready to run with no further intervention. (“Saving” usually means from the web UI, but could also mean uploading a new XML configuration via REST or CLI.)
When a non-administrator saves a template configuration, a check is done whether any contained scripts have been edited from an approved text. (More precisely, whether the requested content has ever been approved before.) If it has not been approved, a request for approval of this script is added to a queue. (A warning is also displayed in the configuration screen UI when the current text of a script is not currently approved.)
An administrator may now go to Manage Jenkins » In-process Script Approval where a list of scripts pending approval will be shown. Assuming nothing dangerous-looking is being requested, just click Approve to let the script be run henceforth.
If you try to run an unapproved script, it will simply fail, typically with a message explaining that it is pending approval. You may retry once the script has been approved. The details of this behavior may vary according to the feature plugin integrating this library.
Waiting for an administrator to approve every change to a script, no matter how seemingly trivial, could be unacceptable in a team spread across timezones or during tight deadlines. As an alternative option, the Script Security system lets Groovy scripts be run without approval so long as they limit themselves to operations considered inherently safe. This limited execution environment is called a sandbox. (Currently no sandbox implementations are available for other languages, so all such scripts must be approved if configured by non-administrators.)
To switch to this mode, simply check the box Use Groovy Sandbox below the Groovy script’s entry field. Sandboxed scripts can be run immediately by anyone. (Even administrators, though the script is subject to the same restrictions regardless of who wrote it.) When the script is run, every method call, object construction, and field access is checked against a whitelist of approved operations. If an unapproved operation is attempted, the script is killed and the corresponding Jenkins feature cannot be used yet.
The Script Security plugin ships with a small default whitelist, and integrating plugins may add operations to that list (typically methods specific to that plugin).
But you are not limited to the default whitelist: every time a script fails before running an operation that is not yet whitelisted, that operation is automatically added to another approval queue. An administrator can go to the same page described above for approval of entire scripts, and see a list of pending operation approvals. If Approve is clicked next to the signature of an operation, it is immediately added to the whitelist and available for sandboxed scripts.
Most signatures be of the form method class.Name methodName arg1Type arg2Type…,
indicating a Java method call with a specific “receiver” class (this), method name, and
list of argument (or parameter) types. (The most general signature of an attempted method
call will be offered for approval, even when the actual object it was to be called on was
of a more specific type overriding that method.) You may also see staticMethod for
static (class) methods, new for constructors, and field for field accesses (get or
set).
Administrators in security-sensitive environments should carefully consider which
operations to whitelist. Operations which change state of persisted objects (such as
Jenkins jobs) should generally be denied. Most getSomething methods are harmless.
Be aware however that even some “getter” methods are designed to check specific
permissions (using an ACL: access control list), whereas scripts are often run by a system
pseudo-user to whom all permissions are granted. So for example method hudson.model.AbstractItem getParent (which obtains the folder or Jenkins root containing
a job) is in and of itself harmless, but the possible follow-up call method hudson.model.ItemGroup getItems (which lists jobs by name within a folder) checks
Job/Read. This second call would be dangerous to whitelist unconditionally, since it would
mean that a user who is granted Job/Create in a folder would be able to read at least some
information from any jobs in that folder, even those which are supposed to be hidden
according to a project-based authorization strategy; it would suffice to create a job in
the folder which includes a Groovy script like this (details would vary according to the
integrating plugin):
println("I sniffed ${thisjob.getParent().getItems()}!");
When run, the script output would display at least the names of supposedly secret
projects. An administrator may instead click Approve assuming permission check for
getItems; this will permit the call when run as an actual user (if the integrating
plugin ever does so), while forbidding it when run as the system user (which is more
typical). In this case, getItems is actually implemented to return only those jobs which
the current user has access to, so if run in the former case (as a specific user), the
description will show just those jobs they could see anyway. This more advanced button is
shown only for method calls (and constructors), and should be used only where you know
that Jenkins is doing a permission check.
For a typical Groovy integration, in which you offer the user the option of using either
script approval or the sandbox, change your describable’s String-valued script field into
a SecureGroovyScript field. In your constructor, before storing the value, call
configuringWithKeyItem (if there could only be one such script per top-level item) or
configuringWithNonKeyItem (if there might be several). The configuration form should use
<f:property field="…"/> to pick up the script and sandbox configuration. When you want
to run the script, just call evaluate.
(For compatibility with old data, pick a different field name and deprecate the original.
Then you can define a readResolve method which sets the new field to a SecureGroovyScript
with the sandbox off, calls configuring(ApprovalContext.create()) on it to notify the
system that an unapproved script has been loaded, and unsets the old field.)
To be used if you need more control than SecureGroovyScript offers:
Introduce a boolean sandbox field into your configuration.
When unset, you need to call ScriptApproval.configuring in the @DataBoundConstructor.
Use ApprovalContext.withCurrentUser, and also withItemAsKey where applicable (when
there is just one script per job); otherwise at least withItem where applicable, and/or
withKey when you can uniquely identify this usage from the context
(StaplerRequest.findAncestorObject is helpful here). This lets the system know a
(possibly) new script has been configured by a particular person. You will also need a
readResolve that calls configuring to notify the system when a configurable with script
has been loaded from disk (and thus the configurer is unknown). Call
ScriptApproval.using when the script is run, and catch UnapprovedUsageException if
necessary. The descriptor should use form validation on the script field and call
ScriptApproval.checking (generally your descriptor should already be doing at least a
syntax check on this field).
When the sandbox field is set, you need merely set up the Groovy shell with
GroovySandbox.createSecureCompilerConfiguration and then call GroovySandbox.run; be
prepared to catch RejectedAccessException and call ScriptApproval.accessRejected.
To preapprove some particular method calls, simply annotate them with @Whitelisted if in
your plugin; otherwise you can register (with @Extension) a ProxyWhitelist delegating to
StaticWhitelist.from and loading a text file listing whitelisted methods.
When constructing a GroovyShell to evaluate a script, or calling
ecureGroovyScript.evaluate, you must pass a ClassLoader which represents the effective
classpath for the script. You could use the loader of Jenkins core, or your plugin, or
Jenkins.getInstance().getPluginManager().uberClassLoader.
Whatever you choose, do not allow an unprivileged user to add arbitrary classpath entries
by making a URLClassLoader! This would make it trivial to bypass all security when using
the sandbox. (A user need merely make this or another job archive a JAR containing some
class with a static method marked @Whitelisted and doing whatever they like, then call
the method from their script.) No attack has yet been demonstrated when using whole-script
approval—a URLClassLoader with normal parent-first delegation would not permit trivial
masking of innocent-looking APIs by compromised versions—but it is likely that some clever
use of META-INF/services/org.codehaus.groovy.transform.ASTTransformation or similar
could cause an otherwise safe script to behave in an unexpected and unauthorized manner.
JENKINS-22834 suggests a safe
standard alternative.
When writing tests for plug-ins that use the Script Security Plugin you may encounter some errors in your tests.
If your tests call, direct or indirectly, the ScriptApproval.get() method, then your
unit tests must use JenkinsRule so that Jenkins.getInstance() does not return null. It
is likely that tests that were working now start to fail if you are not using the sandbox.
It occurs because they are being enqueued for approval. In case you need to execute
scripts regardless of approvals, ScriptApproval.get().preapprove(script, GroovyLanguage.get()) will ensure that all configured scripts are approved. Alternately,
you can have your tests run scripts using the sandbox. In this case you may need to
whitelist methods used by your tests -- either generally for real users, or using a
@TestExtension to have a whitelist just for tests.
See the changelog