JS Forensics

JS Forensics is a static-analysis toolkit for JavaScript that is minified, obfuscated, or simply unfamiliar. It parses a file, reports what is in it, and applies a small set of safe, traceable clean-up transformations.

It never executes the file it analyzes and never contacts a URL it finds.

Status: v0.1.0, alpha. The analyses are heuristic and the transformations are deliberately conservative. See Known limitations before relying on a result.

What it does

• Parses modern JavaScript with tree-sitter and reports syntax errors with line and column.
• Counts functions, declarations, identifiers, literals, calls, loops, and conditionals from the syntax tree.
• Detects twelve documented obfuscation indicators and shows the evidence for each one.
• Extracts string literals and sorts them into URLs, domains, API paths, HTTP header names, cookie strings, HTML, JSON, and everything else.
• Decodes escape-heavy literals and constant String.fromCharCode, String.fromCodePoint, and atob calls without running them.
• Folds constant expressions such as "a" + "b", 0x3e8 * 0x3c, and !0.
• Reports hex-style (_0x1a2b) and very short identifier names.
• Lists references to browser APIs (cookies, storage, network, canvas, WebGL, audio, WebRTC) and whether each is read, written, called, or constructed.
• Lists potential network endpoints with the HTTP method when the source determines it.
• Writes a terminal summary, JSON, or a standalone HTML report.

Installation

Python 3.10 to 3.14 is supported.

$ pip install js-forensics
$ js-forensics --version
js-forensics 0.1.0

To install from source instead:

$ git clone https://github.com/stackadnan/js-forensics.git
$ cd js-forensics
$ python -m venv .venv && source .venv/bin/activate
$ pip install .

The Quick start examples use examples/sample.js from the repository.

Quick start

The repository ships a small obfuscated file to try things on.

$ js-forensics analyze examples/sample.js
$ js-forensics analyze examples/sample.js --format json --output report.json
$ js-forensics strings examples/sample.js
$ js-forensics endpoints examples/sample.js
$ js-forensics deobfuscate examples/sample.js --output-dir output/
$ js-forensics report examples/sample.js --format html --output report.html

analyze prints a summary. Part of the output for the sample:

Obfuscation indicators (heuristic score 45/100)
The score adds fixed weights for the indicators below. It is not a probability and not a verdict on
what the code does.
[+15] Escape-encoded string literals
    -  4 of 8 string literals spell printable ASCII text with \x or \u escapes.
    1:109  Escape-encoded literal.  "\x74\x6f\x6b\x65\x6e"
[+15] Hex-style identifier names
    -  6 of 6 declared names look like _0x1a2b.
[+10] Decoder calls with constant arguments
    1:167  atob call with constant arguments.  atob("L3YxL3Nlc3Npb24=")
[+5] Minified layout

Decoded values
Technique            At     Value
escape-sequences     1:14   https://api.example.com
escape-sequences     1:109  token
atob                 1:167  /v1/session
String.fromCharCode  1:271  POST

Browser APIs referenced
API    Category  Usage  Count  Confidence  First at
fetch  network   call   1      high        1:335

Potential network endpoints
Method   URL                      Source               Resolved  At
UNKNOWN  /v1/session              constant-expression  static    1:167
UNKNOWN  https://api.example.com  string-literal       static    1:14

Unresolved call  At     Reason
fetch            1:335  the URL argument is not a static string

deobfuscate turns the one-line sample into this, and records every change in output/transformations.json:

var _0x3e1f=["https://api.example.com","token"];
function _0x2b7c(_0x1d){
  var _0x4a="/v1/session";
  if(true){
    return _0x3e1f[0x0]+_0x4a+"?"+_0x3e1f[0x1]+"="+_0x1d
  }
  return false
}
var _0x51="POST",_0x77=60000;
fetch(_0x2b7c("abc"),{method:_0x51,keepalive:!![]});
setTimeout(function(){
  console["log"]("done",_0x77)
},1000);

The names stay as they were and the array lookups are not inlined. That is the current boundary of what the tool does, not an oversight.

Commands

Options shared by the analysis commands: --output/-o, --tolerant (analyze a file despite syntax errors and mark the results as partial), --no-redact, and --max-bytes (default 2 MiB).

Exit codes: 0 success, 2 invalid command line, 3 input file missing, unreadable, too large, or not UTF-8, 4 JavaScript syntax error, 5 output could not be written. 1 is an unexpected internal error.

docs/usage.md is the full reference, including the counting rules and every indicator's threshold. docs/output-schema.md describes the JSON.

Supported analysis techniques

Deobfuscation pipeline

Four stages run in order, each on the output of the one before:

1. format-statements puts each statement on its own line. It only changes whitespace between statements.
2. simplify-string-escapes rewrites "\x68\x69" as "hi".
3. decode-static-calls replaces decoder calls on constants with the string they return.
4. fold-constants replaces constant operator expressions with their value.

Transformations edit exact source ranges; the program is not regenerated from its syntax tree, so untouched code stays byte-for-byte as it was. After every stage the result is parsed again, and a stage whose output does not parse is discarded and reported as reverted. Each stage's output is saved under stages/ so you can diff them.

An expression that cannot be resolved safely is left unchanged. Decoder calls and all-constant expressions that were skipped are listed with the reason, for example 1 / 0 ("/ by zero yields Infinity or NaN").

Security boundaries

Every input is treated as hostile.

• The analyzed file is read and parsed. It is never evaluated, imported, or passed to a JavaScript runtime.
• No network access happens. URLs are extracted as text and nothing is resolved or fetched. The test suite fails if a socket is opened.
• The original file is never modified. Outputs go to a separate path and a write that would land on the input is refused.
• Input size is limited (2 MiB by default), nesting depth is bounded in the evaluator, and a transformation that would grow a file more than fourfold is reverted.
• HTML reports contain no script, carry a default-src 'none' Content-Security-Policy, and escape all text taken from the file. Extracted URLs are shown as text, not links.
• Terminal and HTML output escape control characters, so a file cannot inject ANSI sequences or bidirectional overrides into what you read.
• Likely credentials (JWTs, common API-key formats, bearer tokens, URL passwords, secret-looking query parameters and cookie values, and literals assigned to names like password or apiKey) are masked in reports. This is pattern matching and will miss some. --no-redact turns it off. The deobfuscated JavaScript itself is never redacted.
• A finding describes source text. "References document.cookie" does not mean a cookie was read, and a listed endpoint does not mean a request is made.

Known limitations

• No data-flow analysis. Values are resolved only where they are constant at the point of use, so string-array lookups (_0x1f2a(0x1b)), strings built across statements, and values held in variables are not recovered.
• No renaming, no dead-code removal, and no control-flow unflattening.
• Code inside eval, Function, and string timers is not analyzed.
• Constant folding skips anything that needs type coercion beyond string-plus-number, such as "5" * 2, [] + [], +"3", logical and conditional operators, and string comparison with <.
• Template literals are extracted but never rewritten.
• The formatter adds line breaks between statements only. Long expressions and comma sequences stay on one line.
• Decoder calls are resolved on the assumption that String and atob are the built-ins. They are skipped when the file declares or assigns those names or uses with, but runtime monkey-patching cannot be seen.
• Identifier counts are by name and ignore scope.
• The parser is tree-sitter's JavaScript grammar. It also accepts JSX and does not report every early error an engine would. TypeScript is not supported.
• Only UTF-8 input is accepted.
• Memory use is high. On the development machine, a full analysis with transformations of 1 MB of dense minified source took about 9 seconds and peaked near 370 MB. This is why the default input limit is 2 MiB.

Contributing

See CONTRIBUTING.md for setup, the checks to run, and the rules for adding detectors and transformations. docs/architecture.md explains how the code is laid out.

License

MIT. See LICENSE.

Security problems should be reported privately; see SECURITY.md.