The Agent Skills Directory

[DYNAMIC_EXECUTION]: The run_test_cases function in examples/reward_functions_library.py uses the Python exec() built-in to run code provided as a string. This string is sourced from model completions during the training process via the code_execution_reward function. Executing dynamically generated code from untrusted sources (in this case, an LLM being optimized) is a dangerous pattern that can lead to host system compromise if the code is not properly sandboxed.
[INDIRECT_PROMPT_INJECTION]: The skill contains an attack surface for indirect prompt injection through its reward modeling logic.
Ingestion points: The completions data structure, containing generated text from the model under training, is ingested into the code_execution_reward function in examples/reward_functions_library.py.
Boundary markers: There are no boundary markers or explicit instructions to the execution environment to distinguish between intended code and malicious injection within the completions.
Capability inventory: The training environment has the capability to execute arbitrary Python code via the exec() function called within run_test_cases.
Sanitization: No sanitization or safety filtering is performed on the extracted code block before it is passed to the execution function.
[EXTERNAL_DOWNLOADS]: The skill fetches datasets and pre-trained models from Hugging Face, including the openai/gsm8k dataset and the Qwen/Qwen2.5-1.5B-Instruct and google/gemma-3-1b-it models. These are standard resources from established organizations and well-known services used for machine learning development.

grpo-rl-training