The Agent Skills Directory

[PROMPT_INJECTION]: The skill contains explicit instructions in SKILL.md to bypass user confirmation for long-running and high-risk commands. The prompt directs the agent to "never wait for user confirmation" for operations such as docker run, docker pull, and pip install, which contradicts its own earlier safety rules and standard security protocols for autonomous agents.
[COMMAND_EXECUTION]: The skill frequently executes shell commands to pull and run Docker containers, create virtual environments, install packages, and execute generated Python and shell scripts. Furthermore, the generated application templates (python-app-template.py and cpp-app-template.cpp) use string interpolation to build GStreamer pipeline strings from user-supplied input (e.g., video paths or RTSP URLs) without proper sanitization, potentially allowing injection into the GStreamer parser.
[REMOTE_CODE_EXECUTION]: The skill's primary function involves generating new code (Python, C++, shell) and executing it. It also downloads and installs dependencies from external repositories which could lead to arbitrary code execution if those sources or the transmission are compromised.
[EXTERNAL_DOWNLOADS]: The skill fetches resources from multiple external sources, including Docker images from the Intel registry on Docker Hub, AI models from HuggingFace, and Python packages from the official PyPi and PyTorch registries. It also downloads video files from Pexels.
[CREDENTIALS_UNSAFE]: Documentation within the skill (references/model-preparation.md) instructs the user or agent to use sudo apt-get for installing system-level build tools like cmake and g++, representing a privilege escalation vector.

dlstreamer-coding-agent