The Agent Skills Directory

[DYNAMIC_EXECUTION]: The run_test_cases function in examples/reward_functions_library.py uses the Python exec() function to execute code strings. These strings are sourced from code_execution_reward, which processes untrusted completions from the model being trained. Executing dynamically generated code without isolation is a significant security risk.\n- [INDIRECT_PROMPT_INJECTION]: The skill is designed to feed model-generated completions directly into an execution environment.\n
Ingestion points: The completions parameter in code_execution_reward (in examples/reward_functions_library.py).\n
Boundary markers: None are present to prevent the execution of malicious logic within the extracted code.\n
Capability inventory: The use of exec() provides comprehensive access to Python's execution capabilities.\n
Sanitization: The extract_code_block function identifies code sections via regex but does not perform any security validation or content filtering.\n- [COMMAND_EXECUTION]: Arbitrary Python code execution via exec() allows the training process to perform system-level operations, including running shell commands or modifying the filesystem, if the model generates such code.\n- [METADATA_POISONING]: There is an inconsistency in authorship; the skill's files claim it is maintained by 'Orchestra Research' and link to 'orchestra.com', whereas the actual skill author is identified as 'ovachiever'. This can mislead users regarding the skill's origin.

grpo-rl-training