The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes system tools such as dpkg-query, ldd, cmake, and nvcc, along with SDK sample binaries. These executions follow security best practices by using list-based arguments to avoid shell interpolation, running under isolated Python (python3 -I), and employing a restricted environment (SAFE_ENV) to prevent environment-based attacks.
[DYNAMIC_EXECUTION]: The script scripts/query_native_sample_reports.py performs runtime compilation of official SDK samples using cmake and g++. This is a core functionality for hardware verification. The skill mitigates potential risks by verifying the integrity of the source files via the system package manager (dpkg) and re-hashing the resulting binaries before they are executed.
[INDIRECT_PROMPT_INJECTION]: The skill ingests data from external sources including device tree files (/proc/device-tree/model) and command outputs. It implements strong boundary markers and explicit instructions to reject specific request types, such as DRM-protected content or quality metrics. All ingested data is processed using strict regular expressions and JSON schema validation, minimizing the risk of prompt injection from data sources.
[SAFE]: The skill implements extensive self-authentication logic in scripts/_capability_contract.py. It verifies the hashes and versions of the PyNvVideoCodec package against the installed wheel's RECORD file and ensures the execution environment is untainted by environment variables before proceeding with sensitive capability queries. It also uses secure file I/O operations (e.g., O_NOFOLLOW and O_EXCL) to prevent symlink-related vulnerabilities.

jetson-video-capability