The Agent Skills Directory

[PRIVILEGE_ESCALATION]: The skill's documentation in references/model-preparation.md includes instructions for using sudo apt-get install to install build tools on the host. Given the skill's instructions to perform setup tasks autonomously, this introduces a risk of unauthorized privilege escalation.
[DYNAMIC_EXECUTION]: The skill dynamically generates a Python script (export_models.py) from a template and executes it to handle model conversion tasks. This runtime generation and execution of code increases the risk of arbitrary code execution.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted HTML content from external sources (Pexels) to resolve video URLs (SKILL.md Step 2b). 1. Ingestion points: Scraped HTML from Pexels. 2. Boundary markers: Absent. 3. Capability inventory: write and command tools, including docker run and subprocess execution. 4. Sanitization: Absent. This ingestion of untrusted data into the agent's context creates a vulnerability surface for indirect prompt injection.
[EXTERNAL_DOWNLOADS]: The skill performs automated downloads from various remote sources, including Docker images (intel/dlstreamer), Python packages (pytorch.org), and AI models from Hugging Face. The combination of automated downloads and autonomous execution presents a significant attack surface.
[COMMAND_EXECUTION]: The agent uses the command tool to execute shell operations, including docker run with access to host devices (/dev/dri, /dev/accel), which grants the agent significant control over the host hardware and environment.

dlstreamer-coding-agent