The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill instructs the user to download and install code from a third-party GitHub repository not associated with a known trusted vendor.
Evidence: git clone https://github.com/moojink/robocasa-cosmos-policy.git and "robocasa @ git+https://github.com/moojink/robocasa-cosmos-policy.git" in SKILL.md.
[REMOTE_CODE_EXECUTION]: The skill downloads code from external sources and provides instructions to execute it within the user's environment.
Evidence: The installation of the robocasa-cosmos-policy fork followed by uv pip install -e and execution via uv run in SKILL.md.
[COMMAND_EXECUTION]: The instructions include numerous shell commands for environment configuration, dependency syncing, and running evaluation modules.
Evidence: Multiple uv run, srun, and sbatch commands across SKILL.md, references/libero-commands.md, and references/robocasa-commands.md.
[DYNAMIC_EXECUTION]: The skill includes Python snippets that dynamically generate and write configuration files to the user's home directory.
Evidence: Python code in the 'Common issues' section of SKILL.md that writes a config.yaml to ~/.libero/.
[INDIRECT_PROMPT_INJECTION]: The skill defines a workflow for a Vision-Language-Action (VLA) model that processes external visual observations and language instructions, creating an attack surface for indirect prompt injection.
Ingestion points: Visual observations from simulation environments (LIBERO, RoboCasa) and language instruction strings passed to the evaluation modules.
Boundary markers: None identified in the provided instructions or scripts to delimit or sanitize these inputs.
Capability inventory: The skill has capabilities to write logs to the file system, execute subprocesses via uv run, and access GPU resources.
Sanitization: No evidence of sanitization or filtering for the visual or language inputs processed by the model.

evaluating-cosmos-policy