The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses Python's dynamic module loading capabilities to execute arbitrary code from files. Specifically, scripts/run_numerical_leak_check.py and examples/leak_check_adapter.py utilize the importlib.util and exec_module functions to run Python scripts passed via configuration or command-line arguments.
Evidence: In scripts/run_numerical_leak_check.py, the load_adapter function (lines 40-51) calls spec.loader.exec_module(module) on a user-provided file path.
Evidence: In examples/leak_check_adapter.py, the _load_module function (lines 90-100) similarly uses spec.loader.exec_module(module) to run external factor computation code.
[COMMAND_EXECUTION]: The main runner utilizes multiprocessing to execute checks, which involves spawning subprocesses to run the dynamic calculations.
Evidence: scripts/run_numerical_leak_check.py (lines 512-516) uses ProcessPoolExecutor to parallelize the execution of cases.
[REMOTE_CODE_EXECUTION]: While the skill does not fetch code from the network, it creates an indirect prompt injection surface (Category 8) by ingesting 'factor files' or 'adapters' that the agent is instructed to create or identify. If an attacker provides malicious Python code disguised as a financial factor, the tool will execute it with the privileges of the local environment.
Ingestion points: load_adapter in scripts/run_numerical_leak_check.py and _load_module in examples/leak_check_adapter.py.
Capability inventory: Dynamic code execution via exec_module and file system read/write access.
Sanitization: No validation or sandboxing is performed on the loaded Python modules before execution.

numerical-leak-check