The Agent Skills Directory

[SAFE]: The skill does not contain any detected malicious patterns. It is a legitimate integration for the VideoDB platform.
[PROMPT_INJECTION]: No evidence of malicious prompt injection or override instructions. Prompt parameters within the SDK are used for intended generative and analytical tasks (e.g., scene description, image generation).
[DATA_EXFILTRATION]: No unauthorized data exfiltration detected. Network operations are restricted to official VideoDB service endpoints (videodb.io). Sensitive information such as VIDEO_DB_API_KEY is correctly handled using python-dotenv or environment variables.
[REMOTE_CODE_EXECUTION]: The skill utilizes the official videodb Python SDK. While the documentation mentions a local recorder binary for macOS desktop capture, this is a native component of the SDK's core functionality and is distributed through standard package management channels.
[COMMAND_EXECUTION]: All platform interactions are performed through Python script execution using the SDK. No instances of unsafe shell command construction or dynamic execution of untrusted input were found.
[PROMPT_INJECTION]: (Category 8: Indirect) The skill provides a surface for indirect prompt injection as it ingests untrusted data from video transcripts and screen content for LLM-based summarization. However, the risk is mitigated by the scope of the skill's capabilities and standard platform guardrails.
Ingestion points: Transcripts fetched via video.get_transcript_text() and screen activity descriptions in ws_listener.py.
Boundary markers: Missing in example interpolation code (f"...{transcript_text}").
Capability inventory: Subprocess execution for SDK tasks; file reading/writing in local state directories.
Sanitization: None explicitly demonstrated for text interpolation.
Note: This surface is inherent to the functional purpose of the tool and is assessed as SAFE in the context of this skill's implementation.

videodb