The Agent Skills Directory

[DYNAMIC_EXECUTION]: The documentation describes a 'custom' RL loss type that utilizes an 'import_path' to load and execute external Python code at runtime. This dynamic loading mechanism from computed paths allows for arbitrary code execution if an attacker can control the configuration file or CLI arguments.
[INDIRECT_PROMPT_INJECTION]: The skill describes a system that ingests complex structured data from TOML files and JSON strings via the CLI. This creates a surface for indirect injection where malicious instructions or configurations could be embedded in files the agent is instructed to read.
Ingestion points: TOML configuration files (loaded via the '@' symbol) and CLI arguments (including JSON strings for lists and dicts).
Boundary markers: No delimiters or instructions to ignore embedded commands are specified for the configuration values.
Capability inventory: The skill demonstrates the ability to execute the 'rl' command with extensive parameters, including model compilation, external endpoint definition, and custom code loading.
Sanitization: No sanitization or validation of the input configuration content is described beyond Pydantic type validation.
[COMMAND_EXECUTION]: The instructions provide numerous examples of executing the 'rl' tool via 'uv run'. These commands involve deep integration with the system, including file writes to output directories and interactions with local repositories.

configs