The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill defines a framework where autonomous agents ingest data from external environments. This represents a theoretical surface for indirect prompt injection if environmental observations are maliciously controlled and processed by the agent's logic.
Ingestion points: The step and reset methods in scripts/custom_env_template.py and the training loop in scripts/train_rl_agent.py.
Boundary markers: The provided templates do not implement specific delimiters or instructions to ignore embedded commands in observation data.
Capability inventory: The skill is configured to use tools for file operations (Read, Write, Edit) and command execution (Bash) for model persistence and environment setup.
Sanitization: No specific sanitization or filtering logic is provided for observation data, which is typical for numeric reinforcement learning observations.
[COMMAND_EXECUTION]: The skill includes instructions and templates that use standard shell commands for package installation (uv pip) and basic system checks. These commands target well-known, official reinforcement learning libraries and do not exhibit signs of command injection or malicious intent.

stable-baselines3