The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes local diagnostic scripts and authenticated NVIDIA SDK samples (AppEncCuda, AppDec) to query hardware capabilities. These operations are performed using subprocess.run and subprocess.Popen with fixed, known arguments. The AST detector's flag on subprocess.run in query_native_sample_reports.py is a false positive in context, as the script uses a restricted SAFE_ENV and fixed binary paths resolved from the package manager (dpkg).
[REMOTE_CODE_EXECUTION]: No remote code execution was detected. All scripts are local to the skill, and the skill specifically forbids downloading, installing, or repairing packages, delegating those tasks to a separate setup skill if needed.
[DATA_EXFILTRATION]: No data exfiltration patterns were found. The skill does not perform network operations (no curl, wget, or requests) and results are printed to stdout or written to local files specified by the user. It enforces strict isolated Python usage (python3 -I) to prevent environment injection.
[CREDENTIALS_UNSAFE]: No hardcoded credentials or sensitive file accesses were detected. The skill accesses /proc/device-tree/model and /proc/device-tree/compatible to identify the hardware model, which is a standard diagnostic practice on Jetson platforms.
[PROMPT_INJECTION]: The SKILL.md contains strict instructional boundaries, particularly regarding DRM-protected content (Netflix, Widevine), explicitly instructing the agent not to attempt bypasses, describe DRM certification, or recommend workarounds. This acts as a safety guardrail rather than an injection vector.
[DYNAMIC_EXECUTION]: The skill uses importlib.metadata and RECORD file validation to authenticate that the PyNvVideoCodec library being queried is an official, unmodified wheel. It implements a 'fail-closed' mechanism if binary identities or environment hashes drift during execution.

jetson-video-capability