The Agent Skills Directory

[PROMPT_INJECTION]: The skill is vulnerable to Indirect Prompt Injection (Category 8) in the transcript summarization process.
Ingestion points: Raw transcript text is fetched from YouTube via an Apify actor in scripts/scrape_youtube_outliers.py and scripts/update_transcripts.py.
Boundary markers: None. The transcript is directly interpolated into the prompt string using f-strings: f"""... Transcript: {text[:100000]} ...""".
Capability inventory: The script can execute local system commands via subprocess, perform network requests to Apify and Anthropic APIs, and write data to Google Sheets.
Sanitization: No escaping or validation is performed on the transcript content before it is sent to the LLM. An attacker could embed instructions in a video transcript to manipulate the resulting summary or attempt to influence the agent's behavior.
[COMMAND_EXECUTION]: The script scripts/scrape_youtube_outliers.py executes the external binary yt-dlp using subprocess.run.
Evidence: Multiple calls to run_ytdlp(cmd) where cmd includes the yt-dlp command and user-provided or scraped arguments (keywords, channel URLs).
Mitigation: The script uses the list format for subprocess.run which prevents shell-level command injection, but the skill depends on the presence and integrity of an external executable not managed by the skill itself.

youtube-outliers