The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes a local Python script scripts/recipes/recipe_model.py using the isolated mode flag (python3 -I). This prevents the execution from being influenced by user site-packages or environment variables, which is a recommended security practice for local script invocation.
[DATA_EXPOSURE]: The skill processes user-provided JSON files (intent, recipe, environment) and its own internal catalog. The file operations are confined to these specified paths, and there is no evidence of the skill attempting to access sensitive system files, credentials, or private user data.
[EXTERNAL_DOWNLOADS]: The skill references standard industry Python packages, including numpy, pycuda, and torch. These are well-known libraries required for the skill's stated purpose of video encoding and GPU management. No runtime remote code downloads or execution of scripts from untrusted external sources were found.
[COMMAND_EXECUTION]: When generating output files, the script uses the os.O_EXCL flag and sets restrictive file permissions (0o600). This ensures that output files are only accessible by the owner and prevents accidental overwriting of existing files, mitigating potential symlink attacks or unauthorized data access.
[INDIRECT_PROMPT_INJECTION]: The skill has an attack surface for indirect prompt injection as it ingests untrusted JSON data from files like INTENT_JSON and RECIPE_JSON (Cat 8). However, it implements robust sanitization by using a strict_json_load function that rejects duplicate keys and non-finite numbers, and it employs canonical_sha256 hashing to ensure the integrity of the processed recipes. These programmatic boundaries significantly reduce the risk of the agent following malicious instructions embedded in the data.

jetson-video-recipe