JEP 411: sandboxing use case

[Peter Firmstone]

Hi Olivier,

After JEP 421 (deprecation of finalizers) and a JEP is assigned to 
removal of finalizers, it will be possible to instrument java methods 
and intercept their calls.   While finalizers exist, instrumenting 
constructors would allow finalizer attacks to circumvent the permission 
check.

OpenJDK has no intention of developing a new authorization framework / 
api.   Until such time as finalizers are removed, SM is the only option 
currently available for authorization within the JVM.

You can make a JVM process less privileged, or sandbox a JVM within a 
VM, but if the JVM (without SM) is able to generate a students result, 
then student code will also be able to.   I did consider briefly the 
possibility of using two processes, one for the student code (isolated 
in an unprivileged process or VM) and one for the grading code, but I 
think the grading jvm would be susceptible to some form of injection 
attack, as it has to parse untrusted data.

Be sure to disable Serialization, as SM doesn't prevent its use and any 
other forms of potentially dangerous data parsing.

-Djdk.serialFilter=!*,\

-Dcom.sun.jndi.ldap.object.trustSerialData=false,\

Regards,

Peter.


On 12/01/2022 7:49 am, Sean Mullan wrote:
>
>
> On 1/10/22 9:22 AM, Olivier Cailloux wrote:
>> Dear list,
>>
>> I would like to share my use case for currently using the security
>> manager mechanism (SM) in my software. Now that JEP 411 is there, any
>> advice about any currently existing solution for replacement would be
>> welcome, if this is already possible; alternatively, I hope that a
>> replacement for these needs will be available soon.
>
> You may want to consider container technologies. This is mentioned in 
> the last paragraph of the Motivation section of JEP 411.
>
> --Sean
>
>>
>> My software grades student work. It download their code from GitHub,
>> compiles it, runs it, and observe the results (similar to running JUnit
>> tests, but on pluggable code). Their code is then graded automatically
>> depending on the expected versus actual results.
>>
>> I currently use SM to prevent student code to alter the system on which
>> the code runs or have external impact. I don’t want them to read files
>> or send network requests (they usually do not need to do anything like
>> this for the exercices assigned to them). I currently use a simple “no
>> priviledged calls at all” configuration, where everything that can be
>> forbidden by SM is forbidden for their code, as they only need to be
>> able to deal with their own objects and classes from the JDK that
>> operate “taint-free” (as Chapman Flak puts it), such as classical List
>> or Set structures.
>>
>> Though I do not currently need such more advanced feature, I considered
>> as a good bonus that SM allowed me, if I wanted to, to give exercices
>> that also deal with file writing (through telling SM that their code
>> can access a restricted set of files). If any replacement solution
>> could also allow this kind of flexibility, that would be nice.
>>
>> I am aware that their code could implement a denial of service; I am
>> okay to live with this risk as any resulting damage would be low (worst
>> case, just restart the computer). But I’d like to reduce the risk that
>> their code would read or modify files or other aspects of the system it
>> is running on, for example, as the resulting damage could be much
>> higher (such as: alter the way the system works so that the grading of
>> other students, graded next, would be modified; read personal files
>> from the account that is running the grading software and posting their
>> content on the internet; inadvertently delete files on the host
>> system…)
>>
>> I implement code isolation so that one student code does not see or
>> interact with the code of other students classical using class loader
>> mechanisms, for which JEP 411 does not create problems. But I ignore
>> how to prevent file writing, socket opening, or similar stuff, using
>> other means than SM.
>>
>> My needs resemble (but are not identical to) the ones exposed by
>> Chapman Flack in “JEP 411: Missing use-case: user functions in an
>> RDBMS”, https://marc.info/?m=162216583127042. I share the concern of
>> this poster (https://marc.info/?m=162221303911911) that it currently
>> seems that I’d have to come up with various, specialized mechanisms to
>> prevent various kinds of operations (file system access, socket
>> access, …), which seems inelegant and error-prone.
>>
>> Even after reading the insightful article of Ron Pressler, Shallow Java
>> Sandboxes
>> (https://inside.java/2021/04/23/security-and-sandboxing-post-securitymanager/), 
>>
>> it is unclear to me whether I can get rid of SecurityManager with
>> existing Java 17 technology. Any advice would be welcome. If not
>> possible, please consider this use case when thinking about further
>> progress in replacing the security related APIs. (I am quite worried by
>> the wording of JEP 411 Future Work not mentioning this kind of
>> sandboxing need.)
>>
>> Olivier
>>
>>