Skip to content

Commit

Permalink
DRAFT: JEP-200: Switch Remoting/XStream blacklist to a whitelist (#23)
Browse files Browse the repository at this point in the history
* [JENKINS-47736] Draft JEP.

* Changes from proofreading.

* Some notes on testing, as newly required in #24.

* Noting RemoteClassLoader rule.

* Formatting for links

* Replace plus with backtick

* Assign JEP Number: 200
  • Loading branch information
jglick authored and bitwiseman committed Nov 1, 2017
1 parent 9976fa7 commit feb7c55
Showing 1 changed file with 291 additions and 0 deletions.
291 changes: 291 additions & 0 deletions jep/200/README.adoc
@@ -0,0 +1,291 @@
= JEP-0000: Switch Remoting/XStream blacklist to a whitelist
:toc: preamble
:toclevels: 3
ifdef::env-github[]
:tip-caption: :bulb:
:note-caption: :information_source:
:important-caption: :heavy_exclamation_mark:
:caution-caption: :fire:
:warning-caption: :warning:
endif::[]

.Metadata
[cols="2"]
|===
| JEP
| 200

| Title
| Switch Remoting/XStream blacklist to a whitelist

| Sponsor
| https://github.com/jglick[@jglick]

| Status
// Uncomment the appropriate line.
//| Not Submitted :information_source:
| Draft :speech_balloon:
//| Deferred :hourglass:
//| Accepted :ok_hand:
//| Rejected :no_entry:
//| Withdrawn :hand:
//| Final :lock:
//| Replaced :dagger:
//| Active :smile:

| Type
| Standards

| Created
| 2017 Oct 30

| JIRA
| https://issues.jenkins-ci.org/browse/JENKINS-47736[JENKINS-47736]

//
//
// Uncomment if there will be a BDFL delegate for this JEP.
//| BDFL-Delegate
//| :bulb: Link to github user page :bulb:
//
//
// Uncomment if discussion will occur in forum other than jenkinsci-dev@ mailing list.
//| Discussions-To
//| :bulb: Link to where discussion and final status announcement will occur :bulb:
//
//
// Uncomment if this JEP depends on one or more other JEPs.
//| Requires
//| :bulb: JEP-NUMBER, JEP-NUMBER... :bulb:
//
//
// Uncomment and fill if this JEP is rendered obsolete by a later JEP
//| Superseded-By
//| :bulb: JEP-NUMBER :bulb:
//
//
// Uncomment when this JEP status is set to Accepted, Rejected or Withdrawn.
//| Resolution
//| :bulb: Link to relevant post in the jenkinsci-dev@ mailing list archives :bulb:

|===


== Abstract

Since its early days, Jenkins has used two object serialization frameworks:
Java’s built-in API, from the Remoting layer;
and XStream, for persisting configuration and settings.
Both permit any Java class to be loaded merely by specifying its name in the input.

For years, the Jenkins project has received reports of remote code execution (RCE) attacks involving these frameworks.
Typically the attacks involve fairly exotic classes in the Java Platform, or sundry libraries such as Groovy.
The Jenkins CERT has responded to such reports reactively, by blacklisting the affected classes or packages.
That approach has proven unmaintainable, as there is a constant threat of further exploits using unexamined classes.

This proposal switches the blacklist to a whitelist of a few dozen entries:
for Jenkins to process a third-party class, it must be explicitly approved.
Classes defined in Jenkins core or plugins are exempt, as are a few special categories.

In practice it seems that very few plugins actually need to serialize any (third-party) types outside the whitelist.
Many such cases point to dubious design decisions, but to retain compatibility a few such entries are bundled in core.
Plugins or administrators can also expand the whitelist if regressions arise.

== Specification

The main aspect of the change is the introduction of a `ClassFilterImpl` in Jenkins core which supplants the simple blacklist defined in Remoting.
This filter is applied to all Java or XStream deserialization performed by the Jenkins master (as well as XStream _serialization_, to "fail fast").
A crude blacklist extension API in Remoting introduced in
link:https://jenkins.io/security/advisory/2017-04-26/[2017-04-26] is deprecated
in favor of the more general `ClassFilter.setDefault`, as is the system property `hudson.remoting.ClassFilter.DEFAULTS_OVERRIDE_LOCATION`.

The new filter has several levels of rules:

* A `CustomClassFilter` implementation may override any behavior. This is discussed in the compatibility section.
* Classes mentioned in the original blacklist continue to be rejected.
* Java arrays and enums are accepted, as these are special formats.
* Java `Throwable` types are accepted, as they are commonly used in Remoting responses and it is impractical to enumerate every exception that might be thrown.
* A class defined in Jenkins core, Remoting, or a Jenkins plugin is accepted. This does _not_ apply to classes defined in other JARs packaged in `WEB-INF/lib/*.jar`.
* Some classes defined in test utilities, by mocking frameworks, etc., are accepted when running inside tests.
* Classes defined in a `RemoteClassLoader` (so, already sent to this JVM from a presumably trusted source) are accepted.
This is needed at least by CloudBees Jenkins Operations Center, and clients of the `master-to-master-api` plugin if there are any.
* Any other class is accepted if it is explicitly mentioned in a list `whitelisted-classes.txt` bundled in Jenkins core. These consist of:
** Basic Java utility types (roughly half).
** Some other collection types defined in Google Guava.
** A handful of safe types defined in some libraries used by Jenkins core.
** A handful of types used by some common plugins known to otherwise fail: Git, Pipeline, Maven, Job DSL.

The whitelist has been built up by repeatedly running a large array of tests against a patched core and collecting relevant errors.
It is expected that the standard whitelist will be expanded somewhat as field reports arrive of less common code paths needing new entries.

When a class is (first) rejected, a warning is printed to the system log.
This is particularly useful when scanning large automated test results for regressions caused by this change.
Typically the rejection will also trigger an exception which gets displayed in some other manner.

== Motivation

The past few years have seen a flurry of activity by security researchers regarding Java deserialization vulnerabilities.
The `ysoserial` attack library has been created to host standard "gadgets";
Moritz Bechler has
link:https://github.com/mbechler/marshalsec/[published a survey of the field].

While none of the Jenkins CERT team members are experts in this area,
various parties have reported remote code execution (RCE) attacks targeting Jenkins.
In just the past two years, the CERT team has had to issue five security advisories including fixes for deserialization vulnerabilities:
first in
link:https://jenkins.io/security/advisory/2015-11-11/[2015-11-11],
when a new `ClassFilter` blacklist was introduced as a defense; then in
link:https://jenkins.io/security/advisory/2016-02-24/[2016-02-24],
link:https://jenkins.io/security/advisory/2016-11-16/[2016-11-16],
link:https://jenkins.io/security/advisory/2017-02-01/[2017-02-01], and
link:https://jenkins.io/security/advisory/2017-04-26/[2017-04-26].
At this point it is difficult to have any confidence that the ever-growing blacklist in fact covers every dangerous class
bundled in the Java Platform, Jenkins core, or commonly used plugins.
Any newly discovered exploit could be a critical breach in Jenkins security, and it may not be responsibly disclosed.

The exploit in the last (2017-04-26) advisory, like many of the others, was reported against the Jenkins CLI tool.
Since this historically used Jenkins Remoting, it allowed remote attackers—often even with no authentication—to run code inside the Jenkins master.
The fallout from this exploit led the CERT team to deprecate use of Remoting in CLI and switch to a safer protocol:
link:https://gist.github.com/jglick/9721427da892a9b2f75dc5bc09f8e6b3[JENKINS-41745].
Thus Java deserialization exploits are no longer a threat to users of the recommended CLI modes.

Similarly, after 2017-02-01 a potential attack vector involving console notes (markup in Jenkins build logs) was closed:
these must now be signed by a key available only inside Jenkins, and deserialization is only performed after successful signature verification.

However, deserialization is still performed on data an attacker could control in two cases.
Messages sent from an agent to the Jenkins master (unprompted, or responses to requests) are normally passed through a "callable whitelist" as of
link:https://jenkins.io/security/advisory/2014-10-30/[2014-10-30].
This whitelist is only applied _after_ deserializing the message, though, at which point it may be too late.
Since an agent JVM is assumed to be compromisable with a little effort by a rogue build (for example, of a malicious pull request),
the master must apply a filter on incoming classes.

XStream deserialization is also performed when loading job (agent, …) definitions from several REST or CLI commands.
These commands require some authentication and authorization,
but it is worrisome that XStream does not require that a class implement the `Serializable` interface,
so the reserve of potentially exploitable classes is far broader.
Thus any blacklist which hopes to be exhaustive must include many more classes than typical gadgets attempt to use.

(Note: Pipeline builds based on the Groovy CPS engine use yet another serialization framework, JBoss Marshalling, to save state.
This is not considered a security issue since the `program.dat` files are never read from user data.)

== Reasoning

The CERT team could continue to expand the blacklist in response to newly reported vulnerabilities.
This has proven to be a significant maintenance burden, and there is little trust in the result.
Outside security authorities have repeatedly urged the Jenkins team to switch to a whitelist.

Jenkins could theoretically switch to other designs that do not involve Java object deserialization.
In practice this would be wildly incompatible, requiring a rewrite of much of Jenkins core and most plugins.

Every single class used in serial form by Remoting or XStream could be listed.
This would be a gigantic list, however, and would consist mostly of types defined in plugins (thus being antimodular):
it is perfectly common to define callables, settings, or nested "structs" in a plugin for purposes of communication or persistence.
It seems a reasonable compromise to expect that classes defined specifically for use in Jenkins not expose unsafe deserialization behaviors.

In the other direction, it would be possible to reduce the size of the whitelist
by automatically approving any third-party class which does not define a custom deserialization method such as `readResolve`.
(There are some tricky points here involving subclasses, since the Serialization specification allows some inheritance of behaviors.)
This would defend against the most obvious attacks which involve unexpected code execution during deserialization of the exploited class itself.
However, some more subtle gadgets rely on a combination of behaviors:
custom deserialization methods in quite standard classes (usually some kind of collection) which call methods like `equals` or `hashCode` on elements;
and unusual classes which have unsafe implementations of these methods.
Some experimentation was done on this strategy,
but in fact the whitelist size increase needed to handle third-party classes with no deserialization methods is not dramatic,
and this seems well worth the added measure of safety and transparency.

http://openjdk.java.net/jeps/290[JDK Enhancement Proposal (JEP) 290] provides a standard way to apply deserialization filters in Java.
This is not particularly helpful for Jenkins.
There are two kinds of filters in JEP 290: declarative and programmatic.
The programmatic filters would allow the full flexibility that Jenkins’ `ClassFilter` requires.
However, this is only available in Java 9 and later, and anyway we already control the `ObjectInputStream` construction, so it would be functionally equivalent.
(But with no XStream support.)
The declarative filters are available in Java 8, but are too limited
(for example, we cannot automatically approve types defined in Jenkins code);
these have the advantage of applying to any `ObjectInputStream` in the system,
but that is only really helpful when defending against attacks like the `SignedObject` exploit in 2017-04-26,
which was already covered by a blacklist entry (and now a lack of whitelisting as well).

== Backwards Compatibility

There is an obvious risk that some plugins will have a legitimate need to serialize and deserialize third-party types not covered in the whitelist.
In fact it is expected that there will be some such cases;
this is simply the cost of having a tighter security policy.

To ameliorate the risk we can check automated test results against the patched core,
specifically scanning for the term `class-filter` which appears in logs whenever a violation is encountered.
Some runs of `acceptance-test-harness` (ATH) were already performed in this mode.
`plugin-compat-tester` (PCT) was also run against an array of plugins (including those bundled in CloudBees products);
unfortunately the Jenkins project currently has no maintained CI job running PCT against all plugins suggested by the setup wizard.

If new whitelist entries are needed after release, they can be added to core in weekly updates.
Plugins can also contribute their own whitelist (or even blacklist) entries for third-party libraries they bundle,
by creating `META-INF/hudson.remoting.ClassFilter` entries.
(An extension point `CustomClassFilter` is defined allowing _dynamic_ expansions,
but currently not exposed as an API, pending a demonstrated use case.)

Finally, an individual administrator can define site-specific whitelist (or blacklist) entries with a system property `hudson.remoting.ClassFilter`.
This could be useful as an emergency measure, permitting functionality to be restored while awaiting a new plugin release.
(Such a command-line option could be noted as a workaround in a JIRA bug report by someone familiar with the Jenkins security architecture.)

== Security

This proposal is expected to strictly improve Jenkins security,
as the existing blacklist is retained as a fallback unless deliberately overridden.

== Infrastructure Requirements

A new redirect `https://jenkins.io/redirect/class-filter/` will be needed, perhaps pointing to a wiki page.
This permalink is printed to log messages appearing when a whitelist violation is encountered;
in these cases plugin developers or administrators are likely to need instructions on how to proceed.

== Testing

The reference implementation includes test coverage for the essential aspects of the newly added filter:
for example, that an example library class not currently included in the whitelist is rejected under the expected conditions.

A number of core tests had already been added during various advisories as mentioned in the motivation.
When the fallback to the original blacklist is disabled, these continue to pass, indicating that the whitelist alone is a good defense.
(In a few cases, some technical changes had to made to these tests to ensure that they exercised a realistic code path.)

The interesting testing is however driven by scanning ATH and PCT results for failures mentioning certain keywords,
as detailed in the discussion on backwards compatibility.
The broader the set of plugins which can be included in these test runs, the more regressions will be caught early.

For example, a mistake in the `dockerhub-notification` plugin (that would have caused errors under this proposal)
was already detected by an automated test run, and a simple fix proposed and merged.

Testing against this proposal also rediscovered
link:https://issues.jenkins-ci.org/browse/JENKINS-47158[JENKINS-47158],
though sufficient reasonable whitelist entries were added to not cause regressions for Blue Ocean even if that were not fixed.

In several cases, test failures and consequent whitelist additions highlighted poor design decisions in existing code.
For example, as of
link:https://github.com/jenkinsci/git-plugin/pull/497[PR 497]
the `git` plugin does a lot of tricky things with the Eclipse JGit library.
That is true even if you have specified the CLI implementation of Git for use in the build!
In this case, `GitSCM.printCommitMessageToLog` asks the agent to return a `RevCommit` (a JGit type),
which is serialized and deserialized, and then the master calls `getShortMessage()` on that structure.
It would be simpler, faster, and safer to do this processing on the agent and send back a `String`,
but the deceptive ease of Remoting tempts developers to do the wrong thing.
Enforcing a whitelist in the baseline version of Jenkins might help guide them to the simpler solution.

Functional tests (using `JenkinsRule`) which employ mocking frameworks (Mockito / PowerMock)
force the new filter to be disabled, as the changes to class loading prevent normal operation.
Thus any plugin functionality covered only by mock-based tests might quietly regress.
Fortunately these tests generally check only unit functionality to begin with,
and are not likely to be exercising interesting code paths such as settings storage or remote calls to agents.
For similar reasons, certain tests written in Groovy rather than Java prevent normal filter operation and may fail spuriously.

== Reference Implementation

link:https://github.com/jenkinsci/jenkins/pull/3120/files[Jenkins PR 3120] contains the bulk of the change and links to related PRs.

== References

N/A

[IMPORTANT]
====
When moving this JEP from a Draft to "Accepted" or "Final" state,
include links to the pull requests and mailing list discussions which were involved in the process.
====

0 comments on commit feb7c55

Please sign in to comment.