evaluating-cosmos-policy
Pass
Audited by Gen Agent Trust Hub on Nov 1, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSREMOTE_CODE_EXECUTIONINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: Fetches a patched version of the RoboCasa environment and environment configurations from a personal GitHub repository (
moojink/robocasa-cosmos-policy). - [REMOTE_CODE_EXECUTION]: Installs and executes scripts from the external RoboCasa fork, including environment macro setup and asset download utilities.
- [INDIRECT_PROMPT_INJECTION]: The skill displays a vulnerability surface where external model checkpoints and dataset artifacts are ingested and processed by scripts with shell-execution and file-writing capabilities.
- Ingestion points: External model checkpoints (
ckpt_path), dataset statistics (dataset_stats_path), and text embeddings (t5_text_embeddings_path) typically hosted on Hugging Face. - Boundary markers: No delimiters or instructions are used to separate external data content from agent logic.
- Capability inventory: The skill utilizes shell command execution via
uv runandpython -m, and creates local configuration files (e.g.,~/.libero/config.yaml). - Sanitization: No explicit validation or sanitization of the content within model checkpoints or dataset files is performed before processing.
- [COMMAND_EXECUTION]: Performs various shell operations to configure GPU rendering environments (EGL), manage simulation assets, and execute benchmarking loops.
Audit Metadata