The Agent Skills Directory

[PROMPT_INJECTION]: No evidence of malicious instructions, behavioral overrides, or safety filter bypass attempts were found in the documentation or templates.
[DATA_EXFILTRATION]: The skill integrates with well-known logging services like Weights & Biases (wandb) and Neptune for experiment tracking. These are standard practices in machine learning and do not involve unauthorized data exfiltration. No hardcoded credentials or access to sensitive local files were detected.
[EXTERNAL_DOWNLOADS]: The skill references the official PufferLib repository and documentation. It recommends installation via standard package managers (uv pip install pufferlib). These are trusted sources for the skill's functionality.
[COMMAND_EXECUTION]: Shell commands provided in documentation are standard CLI tools for RL training (e.g., puffer train, torchrun) and do not involve dangerous or obfuscated command execution.
[DYNAMIC_EXECUTION]: The skill mentions torch.compile and C extensions for performance optimization. These are standard features in high-performance PyTorch applications and do not present a security risk in this context.
[CREDENTIALS_UNSAFE]: While the training template includes an argument for a Neptune API token (--neptune-token), it is handled as a standard command-line argument for user input rather than being hardcoded into the skill content.

pufferlib