Scripting and security

While Elasticsearch contributors make every effort to prevent scripts from running amok, security is something best done in layers because all software has bugs and it is important to minimize the risk of failure in any security layer. Find below rules of thumb for how to keep Elasticsearch from being a vulnerability.

First and foremost, never run Elasticsearch as the root user as this would allow any successful effort to circumvent the other security layers to do anything on your server. Elasticsearch will refuse to start if it detects that it is running as root but this is so important that it is worth double and triple checking.

Do not expose Elasticsearch directly to users, instead have an application make requests on behalf of users. If this is not possible, have an application to sanitize requests from users. If that is not possible then have some mechanism to track which users did what. Understand that it is quite possible to write a _search that overwhelms Elasticsearch and brings down the cluster. All such searches should be considered bugs and the Elasticsearch contributors make an effort to prevent this but they are still possible.

Do not expose Elasticsearch to the Internet, instead have an application make requests on behalf of the Internet. Do not entertain the thought of having an application "sanitize" requests to Elasticsearch. Understand that it is possible for a sufficiently determined malicious user to write searches that overwhelm the Elasticsearch cluster and bring it down. For example:

Good: * Users type text into a search box and the text is sent directly to a Match Query, Match Phrase Query, Simple Query String Query, or any of the Suggesters. * Running a script with any of the above queries that was written as part of the application development process. * Running a script with params provided by users. * User actions makes documents with a fixed structure.

Bad: * Users can write arbitrary scripts, queries, _search requests. * User actions make documents with structure defined by users.

By default Elasticsearch will run inline, stored, and filesystem scripts for sandboxed languages, namely the scripting language Painless, the template language Mustache, and the expression language Expressions. These ought to be safe to expose to trusted users and to your application servers because they have strong security sandboxes. By default Elasticsearch will only run filesystem scripts for non-sandboxed languages and enabling them is a poor choice because: 1. This drops a layer of security, leaving only Elasticsearch’s builtin security layers. 2. Non-sandboxed scripts have unchecked access to Elasticsearch’s internals and can cause all kinds of trouble if misused.

In addition to user privileges and script sandboxing Elasticsearch uses the Java Security Manager and native security tools as additional layers of security.

As part of its startup sequence Elasticsearch enables the Java Security Manager which limits the actions that can be taken by portions of the code. Painless uses this to limit the actions that generated Painless scripts can take, preventing them from being able to do things like write files and listen to sockets.

Elasticsearch uses seccomp in Linux, Seatbelt in macOS, and ActiveProcessLimit on Windows to prevent Elasticsearch from forking or executing other processes.

Below this we describe the security settings for scripts and how you can change from the defaults described above. You should be very, very careful when allowing more than the defaults. Any extra permissions weakens the total security of the Elasticsearch deployment.

Which scripts Elasticsearch will execute where is controlled by settings starting with scripts.. The simplest settings allow scripts to be enabled or disabled based on where they are stored. For example:

script.inline: false  
script.stored: false  
script.file:   true   

Scripting may also be enabled or disabled in different contexts in the Elasticsearch API. The supported contexts are:

Plugins can also define custom operations that they use scripts for instead of using the generic plugin category. Those operations can be referred to in the following form: ${pluginName}_${operation}.

The following example disables scripting for update and plugin operations, regardless of the script source or language. Scripts can still be executed from sandboxed languages as part of aggregations, search and plugins execution though, as the above defaults still get applied.

script.update: false
script.plugin: false

First, the high-level script settings described above are applied in order (context settings have precedence over source settings). Then fine-grained settings which include the script language take precedence over any high-level settings. They have two forms:

script.engine.{lang}.{inline|file|stored}.{context}: true|false

And

script.engine.{lang}.{inline|file|stored}: true|false

For example:

script.inline: false 
script.stored: false 
script.file:   false 

script.engine.painless.inline:          true 
script.engine.painless.stored.search:   true 
script.engine.painless.stored.aggs:     true 

script.engine.mustache.stored.search:   true 

As mentioned above, Elasticsearch runs with the Java Security Manager enabled by default. The security policy in Elasticsearch locks down the permissions granted to each class to the bare minimum required to operate. The benefit of doing this is that it severely limits the attack vectors available to a hacker.

Restricting permissions is particularly important for non-sandboxed scripting languages like Groovy and Javascript which are designed to do anything that can be done in Java itself, including writing to the file system, opening sockets to remote servers, etc.

Groovy makes an effort to prevent loading classes which do not appear in a hardcoded whitelist that can be found in org.elasticsearch.script.ClassPermission.

In a script, attempting to load a class that does not appear in the whitelist may result in a ClassNotFoundException, for instance this script:

GET _search
{
  "script_fields": {
    "the_hour": {
      "script": "use(java.math.BigInteger); new BigInteger(1)"
    }
  }
}

will return the following exception:

{
  "reason": {
    "type": "script_exception",
    "reason": "failed to run inline script [use(java.math.BigInteger); new BigInteger(1)] using lang [groovy]",
    "caused_by": {
      "type": "no_class_def_found_error",
      "reason": "java/math/BigInteger",
      "caused_by": {
        "type": "class_not_found_exception",
        "reason": "java.math.BigInteger"
      }
    }
  }
}

However, classloader issues may also result in more difficult to interpret exceptions. For instance, this script:

use(groovy.time.TimeCategory); new Date(123456789).format('HH')

Returns the following exception:

{
  "reason": {
    "type": "script_exception",
    "reason": "failed to run inline script [use(groovy.time.TimeCategory); new Date(123456789).format('HH')] using lang [groovy]",
    "caused_by": {
      "type": "missing_property_exception",
      "reason": "No such property: groovy for class: 8d45f5c1a07a1ab5dda953234863e283a7586240"
    }
  }
}

If you encounter issues with the Java Security Manager, you have two options for resolving these issues:

The safest and most secure long term solution is to change the code causing the security issue. We recognise that this may take time to do correctly and so we provide the following two alternatives.

The classloader whitelist can be customised by tweaking the local Java Security Policy either:

Permissions may be granted at the class, package, or global level. For instance:

grant {
    permission org.elasticsearch.script.ClassPermission "java.util.Base64"; // allow class
    permission org.elasticsearch.script.ClassPermission "java.util.*"; // allow package
    permission org.elasticsearch.script.ClassPermission "*"; // allow all (disables filtering basically)
};

Here is an example of how to enable the groovy.time.TimeCategory class:

grant {
    permission org.elasticsearch.script.ClassPermission "java.lang.Class";
    permission org.elasticsearch.script.ClassPermission "groovy.time.TimeCategory";
};

See http://docs.oracle.com/javase/7/docs/technotes/guides/security/PolicyFiles.html for more information.