The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill provides a reward function library (examples/reward_functions_library.py) that includes a run_test_cases function. This function utilizes the Python exec() statement to run code strings directly at runtime.\n- [REMOTE_CODE_EXECUTION]: The code_execution_reward function in examples/reward_functions_library.py extracts code from model-generated completions and passes it to the run_test_cases function. This creates an arbitrary code execution vulnerability where the model being trained could generate and execute malicious code on the host training environment.\n- [INDIRECT_PROMPT_INJECTION]: The skill's architecture is vulnerable to indirect prompt injection because it executes data generated by the language model without adequate isolation or sandboxing.\n
Ingestion points: Model completions are ingested into the code_execution_reward function within the examples/reward_functions_library.py file.\n
Boundary markers: No boundary markers or specialized instructions are present to prevent the model from generating malicious code that could be executed during the reward evaluation phase.\n
Capability inventory: The environment executing the training script has the capability to run arbitrary Python code via the exec() call found in examples/reward_functions_library.py.\n
Sanitization: No sanitization, validation, or sandboxing is implemented in the provided reward function code; although a comment notes the requirement for sandboxing, the functional code provided is unsafe.

grpo-rl-training