This is a proof of concept java virus, which infects jars contained in the same directory as itself, and incorporates a few novel java anti-analysis techniques.
The obvious caveat is that this is what I'm aware of. I did a little research when I started out with Bismuth and couldn't find anything, but I missed Strangebrew, Cheshire's author might well have missed Bismuth, and Neko's author did completely their own thing (but probably won't be willing to answer our questions, even if I knew who to ask).
- Virus wikidot article
- Virus Bulletin Sept 1998 (page 11)
- ftp://static.zedz.net/pub/security/info/textfiles/codebreakers-vx-zine/codebreakers-4/ (Codebreakers VX Zine, 4th release)
Widely credited as the first Java virus, written by an Australian university student, and published in the 4th release of Codebreakers VX Zine, this came with a helpful tutorial introducing readers to the Java classfile format, as well as full source code.
The virus injects its own method into nearby .class files, apparently it was a little buggy, but the overall concept was sound.
- Virus Wikidot article
- Kaspersky summary
- ftp://static.zedz.net/pub/security/info/textfiles/codebreakers-vx-zine/codebreakers-5/ (Codebreakers VX Zine, 5th release)
Comparable mechanism to the later Nekoclient/« Fractureiser », although more verbose, infection is with a stub which loads a classfile from a remote URL. Written by the same author as Strangebrew, though the tutorial is sadly not as detailed.
What's that you say? You don't know anything about java? Don't worry your little head about it - Landing Camel
- Virus Bulletin Dec 2013 (page 15)
Self-replication by changing the entry point in the manifest. Crude, probably a better time than dealing with bytecode.
Self-replication by copying the payload/infector into a target jar, and modifying a constructor to point to it. (You are here)
A self-contained example presented at Montreal's REcon in 2022, Cheshire is a masterpiece, it's some really impressive work, and it's what I'd refer people to if they needed an example of this phenomenon. I can empathise with the author's frustration with stack frame maps, and with spending hours staring at the classfile docs!
The final stage of malware involved in an interesting incident attacking Minecraft developers and players, making good use of the inadequate security around Minecraft mods. The final stage does a lot of interesting things (session stealing was used to good effect in the attack), but includes a replication mechanism, albeit not strictly straightforward self-replication, inserting a small stub into classfile constructors to load a classfile from a URL.
The heavy classfile lifting is done by use of the ObjectWeb asm library, as someone who's spent many days living inside Java's classfile docs, I resent this.
The author of this malware seemingly uploaded an unobfuscated version of it by accident (lol), and the source linked above is a decompilation of that. I'm unsure which specific decompiler was used on this repo (this is quite an important consideration when dealing with Java malware), though I know some of those analysing it are aware of CFR, the decompiler I favour for this work, and the code seems to demonstrate the replication part just fine in any case.
I've written previously about the risk of something like this happening around Minecraft, and I suppose Bismuth is my "I told you so" card, but this isn't a very satisfying feeling. I asked for samples in the espernet channel, and was met with an "interesting" response from the self-appointed responders, I'm also aware of some of the discussions which have taken place between developers/maintainers of launchers and mod platforms; I'm not at all confident that security around Minecraft modding will be meaningfully improved.
I look forward to adding more entries to this section!
- Ensure your jarfiles are read only (easy)
- Sign jarfiles, and require signatures (hard)
Creative Commons BY-NC-SA 4.0
This is the entry point. This returns if it's already been run by one of the many invocations from constructors.
This function iterates through jar files in the same directory, calling infectJar on them.
The jarfile is opened. Any signature files present are stripped (it's more likely they'll be present than mandatory), and this method returns early if any filenames contain the infection signature, which consists of every odd-indexed character being 'w' or greater.
ActualScorpion is streamed into into a .class file in the root of the jar, with a random name complying with the above signature, through the transformSelf
function, which obfuscates the class, and updates it with its new name.
A random selection of classfiles present in the jar are streamed through injectInvoke, which inserts an invocation referencing
ActualScorpion's run, reverting the file if this method returns true, indicating a condition which would prevent classfile verification.
This inserts the relevant references to the obfuscated ActualScorpion's run into the constant pool, and prepends an invokestatic instruction
referencing this into the Code attribute of the first method (invariably <init>, the constructor). This is surrounded by bytecode which
frustrates some analysis tools if the classfile version if low enough that stack frame maps aren't mandatory on branches.
The exceptions table in Code is offset to accomodate the new bytecode, as are the LocalVariableTable and LocalVariableTypeTable attributes,
if present.
If a stack frame map is present, this method returns true, to indicate to infectJar that the file needs to be reverted, as it'll fail verification
without appropriate amendments to the stack frame map. Constructors aren't usually complex enough to require stack frame maps, so it's probably
not worthwhile to add this functionality.
This changes ActualScorpion references to refer to the randomly generated name selected in infectJar, and applies obfuscation throughout.
The constant pool is loaded into a linked hashmap and written out at the end, in order to allow manipulation of strings informed by later
contents of the classfile.
- A second
Codestring is inserted into the constant pool, which<init>references. A tool which expects only the first instance ofCodein the constant pool to be used for a method'sCodeattribute will treat the constructor as an empty method. - Constructor access flags are forced to synthetic, a flag intended to indicate members not present in the source. Some tools omit synthetic members from their output
If the classfile version permits, invokestatic is supplemented with branches. Some tools misinterpret wide branches, which can be exploited by using a wide jump into what would otherwise be dead code. The tools regard it as such, and optimise it out.
; A7 00 09 (+9)
goto first
second:
; B8 00 ?? ??
; The hidden invocation!
invokestatic ActualScorpion run ()V
; A7 00 08 (+8)
goto third
first:
; C8 FF FF FF FA (-6)
goto_w second
third:
- The SourceFile string is set to a zalgoified string, containing
\033(0 \033[42;5;35m, an escape sequence which puts a terminal into graphical mode, with magenta text on a green background. This will be displayed on the terminal if an exception occurs anywhere in this file. - All permitted access flags are set. This serves two functions: Making life harder for tools, and causing confusion to a human using them, given
the relative obscurity of some of these keywords
- Class: final, synthetic
- Field: volatile, transient, synthetic
- Method: final, synchronized, bridge, strictfp, synthetic
- Method and field names (aside from
run,<init>and<clinit>) are set to 65280 randomly selected mostly unprintable characters ('\x01'..'-'). Some tools won't escape these, but most will escape unprintable characters in the\uXXXXform, a potential sixfold increase in member name length, every time the member is referenced. - Method
LocalVariableTableandLocalVariableTypeTableattributes are changed to offer no useful hints as to the names of variables - Method
LineNumberTableattributes, which correspond bytecode to line numbers for debugging and analysis, are made less accurate
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent