The Agent Skills Directory

Prompt Injection (HIGH): The skill creates a significant vulnerability to indirect prompt injection via its real-time multimodal input processing design.\n
Ingestion points: examples/video-agent.py utilizes cv2.VideoCapture to ingest video or camera feeds, while examples/voice-agent.py and templates/liverequest-queue.py ingest live audio chunks directly into the agent context.\n
Boundary markers: Absent. There are no instructions to the model to ignore embedded commands within the multimodal stream, and the data is not delimited to distinguish it from system instructions.\n
Capability inventory: The SKILL.md frontmatter explicitly allows the Bash, Read, and Write tools. This high-privilege toolset creates a path for remote command execution if a multimodal injection successfully influences the model.\n
Sanitization: Absent. Multimodal data is passed directly from capture interfaces (OpenCV, microphone) to the LiveRequestQueue without any filtering, validation, or sanitization layers.- Dynamic Execution (LOW): The script scripts/check-modality-support.py uses __import__ to verify module availability, but since the module names are sourced from a hardcoded dictionary rather than user input, the risk is negligible.- External Downloads (LOW): The skill instructions suggest installing standard packages like google-adk and opencv-python. These are verifiable packages from trusted repositories and do not constitute a high-risk finding.

streaming-patterns