The Agent Skills Directory

[PRIVILEGE_ESCALATION]: The installation steps in SKILL.md require the user to run sudo pip uninstall for several packages, which utilizes elevated system permissions. Additionally, the recommended docker run command uses the --cap-add=SYS_ADMIN flag, granting the container extensive administrative capabilities on the host system.
[INDIRECT_PROMPT_INJECTION]: The references/custom-rewards.md file defines an implementation pattern for code generation rewards that is vulnerable to indirect injection.
Ingestion points: The reward_func logic extracts generated_code directly from model-produced queries in references/custom-rewards.md.
Boundary markers: The examples use simple string manipulation (query.split(prompt)[-1]) to isolate model output, which can be bypassed by adversarial content designed to escape the expected response structure.
Capability inventory: The script demonstrates using subprocess.run to execute the extracted code via pytest and the Python interpreter.
Sanitization: The documentation lacks instructions or examples for sandboxing, resource limiting, or validating the model-generated code before execution.
[DYNAMIC_EXECUTION]: references/custom-rewards.md provides examples for dynamic code execution by writing strings to temporary files using tempfile.NamedTemporaryFile and executing them as Python scripts.
[COMMAND_EXECUTION]: The skill relies on shell commands for environment management and distributed task execution, including ray start, ray job submit, and deepspeed module invocations.

openrlhf-training