If a server can deserialize objects from an untrusted source, it is possible to craft a serialized object that runs arbitrary code on deserialization. Classes in the Clojure jar can be used for this purpose and this may lead to blacklisting the Clojure jar in some environments.
Demonstration of how to create such an object: https://github.com/frohoff/ysoserial/pull/68/files
The serialized class being constructed in the attack is clojure.inspector.proxy$javax.swing.table.AbstractTableModel$ff19274a, but clojure.core.proxy$clojure.lang.APersistentMap$ff19274a and proxy classes created by users are also susceptible. Clojure proxies contain a field (__clojureFnMap) that is a map of method names to proxy functions. Clojure AOT compiles the clojure.inspector and clojure.core.proxy code which generates the classes named above. These classes are then included inside the clojure jar.
The attack constructs an instance of one of these proxy classes and adds a "hashCode" proxy method to the proxy's table. The method is a Clojure function that can run arbitrary code. This instance is then put inside a HashMap and the whole thing is serialized. On deserialization, the HashMap will invoke hashCode() on the proxy object and cause the execution of the arbitrary code.
Note that most uses of proxy will create objects that cannot be serialized anyway, due to the common use of literal function bodies.
Proposed: Modify proxy so that the classes it generates are neither deserializable (to disable the attack) nor serializable (to avoid misleading users). The patch causes proxy classes to have readObject and writeObject methods that just throw an exception, when Serializable is in the inheritance tree of the proxy class. This is technically a breaking change since it is possible to construct a proxy that can be serialized, but such classes are unlikely in the wild and inherently insecure.
Prescreened by: Alex Miller
Background: Raised at https://groups.google.com/d/msg/clojure/WaL3hHzsevI/7zHU-L7LBQAJ and similar to issues reported on Apache Commons. See also:
Patch looks good to me - can you add a test with the harness to the patch too?
Thanks for reviewing it.
I will add a test; it could simply confirm that an attempt to serialize a proxy class fails. Do you think that's sufficient, or should it also include the attempted code execution? I guess the latter could fail for other more subtle reasons, causing the test to pass when perhaps we wouldn't want it to, so I'll go for the simpler test first.
I think the simpler test is sufficient.
Attached patch with fix and test. Note the test for prohibited deserialization is tricky because we also prevent the creation of the data needed for that test. Because of this, the serialized stream is included directly, along with instructions on how to recreate the data if needed in the future.
Alex, I just realized (by looking though this ticket's history) that I unintentionally overwrote your Description update last night. Not sure how that happened, but I apologize. Feel free to change it back or take bits from mine and add it to yours, or whatever else you feel is appropriate.