The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The script scripts/test-signature.py uses importlib.import_module to dynamically load Python modules based on user-provided CLI arguments, allowing for the execution of arbitrary module-level code during initialization.
[REMOTE_CODE_EXECUTION]: Documentation in references/programs.md demonstrates loading compiled programs using dspy.load(..., allow_pickle=True). Deserializing untrusted data with pickle is an unsafe pattern that can lead to arbitrary code execution.
[COMMAND_EXECUTION]: scripts/clear-cache.py performs recursive directory deletion using shutil.rmtree on a path controlled by the --cache-dir argument, which could be exploited to delete arbitrary directories.
[PROMPT_INJECTION]: Template files and documentation (e.g., assets/templates/program-template.py) demonstrate patterns where untrusted external data is passed directly to LLM components without sanitization or boundary markers, creating a surface for indirect prompt injection.
Ingestion points: input_data argument in various module templates within assets/templates/program-template.py.
Boundary markers: None identified; input is interpolated directly into DSPy predictors.
Capability inventory: The skill includes scripts for file system modification (clear-cache.py) and dynamic code loading (test-signature.py).
Sanitization: No validation or escaping is applied to external content before processing.

dspy-core