The Agent Skills Directory

[PROMPT_INJECTION]: The skill instructions in SKILL.md contain an override of safety controls. While a general safety section requires user confirmation for dangerous commands, a specific instruction in Step 5 tells the agent to use 'autonomous execution' and 'never wait for user confirmation' for docker pull, pip install, and model export commands.
[REMOTE_CODE_EXECUTION]: The reference guide for model preparation explicitly encourages the use of the 'trust-remote-code' flag with Hugging Face models. This allows for the execution of arbitrary Python code from the remote model repository during the export process.
[COMMAND_EXECUTION]: The skill procedure requires the automated execution of package managers, container tools, and shell commands (docker, pip, curl) to set up the environment and run the generated applications.
[PRIVILEGE_ESCALATION]: The model preparation documentation instructs the agent to use 'sudo' to install system dependencies (cmake, g++), which grants administrative access to the host system.
[INDIRECT_PROMPT_INJECTION]: The skill presents an attack surface for indirect prompt injection. (1) Ingestion points: Video input URLs and model names provided by the user in Step 0. (2) Boundary markers: The procedure mentions allowlists, but no markers or delimiters are implemented in the generated code. (3) Capability inventory: The skill can write and execute scripts that perform network operations and containerized execution. (4) Sanitization: The generated Python and C++ templates build GStreamer pipeline strings using direct concatenation of user-provided inputs, which are then passed to Gst.parse_launch without sanitization.
[EXTERNAL_DOWNLOADS]: The skill downloads AI models from Hugging Face and software packages from the PyPI and PyTorch registries to fulfill its processing requirements.

dlstreamer-coding-agent