attack-surface.md

Attack Surface Analysis for pandas-dev/pandas

Attack Surface: Deserialization Vulnerabilities via read_pickle()

Description: Using pandas.read_pickle() to deserialize data from untrusted sources can lead to arbitrary code execution. Malicious pickle files can execute code during deserialization.

Pandas Contribution: pandas.read_pickle() is the function that performs deserialization of pickle files, inherently executing code embedded within them.

Example: A web application processes user-uploaded .pkl files using read_pickle(). A malicious .pkl file is uploaded, containing code that executes on the server when processed by pandas, leading to server compromise.

Impact: Critical - Arbitrary code execution, full server compromise.

Risk Severity: Critical

Mitigation Strategies:

• Avoid read_pickle() for untrusted data: Do not use read_pickle() to process data from untrusted sources.
• Use safer formats: Prefer safer, text-based formats like JSON or CSV for untrusted data.
• Sandboxing: If read_pickle() is unavoidable, execute it in a heavily sandboxed environment.

Attack Surface: Code Execution via eval() and query()

Description: Pandas functions utilizing eval() or similar dynamic code execution, especially with user-controlled input, can be exploited for code injection. This includes DataFrame.query() and potentially data input functions with specific engines.

Pandas Contribution: DataFrame.query() directly uses string-based expressions that can execute arbitrary code if crafted maliciously. Certain data reading engines in pandas might also use eval() internally.

Example: A web application uses user-provided input to construct a query string for DataFrame.query(). A malicious user injects code into the query string, which is then executed by pandas, leading to server compromise.

Impact: High - Arbitrary code execution, server compromise.

Risk Severity: High

Mitigation Strategies:

• Avoid eval() and query() with user input: Do not use DataFrame.query() or functions relying on eval() with user-provided input.
• Parameterize queries: Use safer filtering methods without string-based code execution.
• Input sanitization (complex and less reliable): Sanitize user input if dynamic queries are absolutely necessary, but this is error-prone for code injection.

Attack Surface: Path Traversal Vulnerabilities in File Path Handling

Description: Pandas file I/O functions (read_csv, to_csv, etc.) accepting file paths derived from user input without validation can lead to path traversal attacks, allowing access or manipulation of files outside intended directories.

Pandas Contribution: Pandas provides functions that take file paths as arguments for reading and writing data. If these paths are constructed from unsanitized user input, pandas becomes a vector for path traversal.

Example: A web application uses user-provided filenames with pandas.to_csv(). A malicious user provides a filename like ../../../../sensitive_data.csv, potentially overwriting or accessing sensitive files due to insufficient path validation in the application using pandas.

Impact: High - Information disclosure (reading sensitive files), data manipulation (overwriting files).

Risk Severity: High

Mitigation Strategies:

• Strict input validation for file paths: Thoroughly validate and sanitize user-provided file paths before using them with pandas functions.
• Allowlisting: Use allowlists for permitted characters and directory paths.
• Avoid direct user input in path construction: Do not directly concatenate user input into file paths. Use secure path construction methods.