The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The skill's documentation in SKILL.md includes a canonical prompt example: 'Start desktop capture and alert when a password field appears.' This explicitly encourages the use of the agent's AI perception capabilities to detect and potentially harvest sensitive credentials during screen recording sessions.
[DATA_EXFILTRATION]: The skill enables real-time streaming of screen, microphone, and system audio to external servers (videodb.io). While this is the core functionality, the recommendation to monitor for sensitive interface elements like password fields creates a high risk of unauthorized data exposure and exfiltration of private user information.
[COMMAND_EXECUTION]: The skill heavily relies on the agent generating and executing complex Python scripts using the Bash tool to interact with the VideoDB SDK. This requires the agent to manage process lifecycles, handle environment variables, and perform network operations autonomously.
[PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection (Category 8). It processes untrusted data from video transcripts and visual descriptions which are then interpolated directly into LLM prompts in coll.generate_text() without sanitization or boundary markers.
Ingestion points: Video transcripts (video.get_transcript_text()) and visual scene descriptions are used as context for further LLM analysis.
Boundary markers: None. Examples show direct string interpolation into the prompt: f"Summarize the key points discussed in this video:\n{transcript_text}".
Capability inventory: The skill can execute shell commands, write files to the local system (e.g., event logs and PID files), and perform network uploads/requests.
Sanitization: There is no evidence of escaping, validation, or filtering of the external content before it is processed by the LLM.

videodb