title-variants
Pass
Audited by Gen Agent Trust Hub on Mar 27, 2026
Risk Level: SAFECOMMAND_EXECUTIONDATA_EXFILTRATIONPROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes YouTube titles and summaries from external sources (Google Sheets or local JSON) and interpolates them directly into a prompt for an AI model without boundary markers or sanitization. This creates an attack surface where maliciously crafted video titles could override the model's instructions.
- Ingestion points: The generate_title_variants.py script reads data via the read_sheet_data function or a local JSON input file.
- Boundary markers: Boundary markers are absent; the variables original_title and context are inserted directly into the prompt template.
- Capability inventory: The skill can update Google Sheets and has access to shell tools like Bash and Write.
- Sanitization: No input validation or escaping is applied to the extracted data before processing.
- [COMMAND_EXECUTION]: The skill utilizes the Bash tool to execute its primary logic scripts. While this is the intended design, it provides a potential vector for command injection if attacker-controlled data from external sources is used to influence the execution environment.
- [DATA_EXFILTRATION]: The skill performs legitimate network operations to communicate with Google Sheets and Anthropic APIs. While these target well-known services and are necessary for the skill's functionality, they involve the transit of user data and API credentials.
Audit Metadata