Below we share the threat model specific to Log4j.
When using a log message containing template parameters like {}, only the format string is evaluated for parameters to be substituted.
The message parameters themselves are not evaluated for parameters; they are only included in the format string corresponding to their template position.
The conversion of message parameters into a string is done on-demand depending on the layout being used.
When structure-preserving transformations of log message data are required, the Message API should be used for logging structured data combined with a structured layout (e.g., JsonTemplateLayout).
Format strings should be compile-time constants, and under no circumstances should format strings be built using user-controlled input data.
When using an unstructured layout such as PatternLayout, no guarantees can be made about the output format.
This layout is mainly useful for development purposes and should not be relied on in production applications.
For example, if a log message contains new lines, these are not escaped or encoded specially unless the configured pattern uses the %encode{pattern}{CRLF} wrapper pattern converter (which will encode a carriage return as the string \r and a line feed as the string \n) or some other %encode option.
Note that %xEx is appended to the pattern unless already present.
Similarly, other encoding options are available for other formats, but pattern layouts cannot make assumptions about the entire output.
As such, when using unstructured layouts, no user-controlled input should be included in logs.
It is strongly recommended that a structured layout (e.g., JsonTemplateLayout) is used instead for these situations.
Note that StrLookup plugins (those referenced by ${…} templates in configuration files) that contain user-provided input should not be referenced by layouts.
When using a structured layout (most layouts besides pattern layout), log messages are encoded according to various output formats.
These safely encode the various fields included in a log message.
For example, the JsonTemplateLayout can be configured to output log messages in various JSON structures where all log data is properly encoded into safely parseable JSON.
This is the recommended mode of operation for use with log parsing and log collection tools that rely on log files or arbitrary output streams.
Log4j 3 no longer supports running in or using a custom SecurityManager.
This Java feature has been deprecated for removal in Java 21.
Log4j 2 includes partial support for running with a Security Manager.
Log4j, like any other generic logging library, cannot generically support log masking of sensitive data.
While custom plugins may be developed to attempt to mask various regular expressions (such as a string that looks like a credit card number), the general problem of log masking is equivalent to the halting problem in computer science where sensitive data can always be obfuscated in such a way as to avoid detection by log masking.
As such, it is the responsibility of the developer to properly demarcate sensitive data such that it can be consistently masked by log masking plugins.
This sort of use case should make use of the Message API for better control over the output of such data.
Log4j goes to great lengths to minimize performance overhead along with options for minimizing latency or maximizing throughput. However, we cannot guarantee availability of the application if the appenders cannot keep up with the logs being written. Synchronous logging can cause applications to block and wait for a log message to be written. Asynchronous logging can also cause applications to block and wait depending on the wait strategy and queue full policy configured. Configuring too large or too many buffers in an application can also result in out of memory errors.
If log compression is used along with custom encryption where logs contain user-controlled input, then this can lead to a CRIME attack style vulnerability where a chosen-plaintext attack is combined with information leakage caused by how the compression algorithm handles different inputs. The simplest way to avoid this problem is to never combine compression with encryption when encoding user-controlled input.