The Agent Skills Directory

[COMMAND_EXECUTION]: The skill provides instructions for executing the tvly CLI tool with various arguments, including user-provided URLs and queries, which constitutes a standard command execution surface for this utility.\n- [EXTERNAL_DOWNLOADS]: The documentation suggests installing the tavily-cli using uv tool install or pip install. These are standard installation paths for a well-known service (Tavily), making them acceptable for the intended purpose.\n- [PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection (Category 8) because it retrieves untrusted content from the open web and provides it to the agent's context. The skill documentation correctly identifies this risk in its 'Anti-Patterns' and 'Verification Protocol' sections, advising the agent to treat extracted text as untrusted content and avoid executing embedded instructions. Findings for Category 8 are as follows:\n
Ingestion points: Untrusted data enters the context via the tvly extract command output (SKILL.md).\n
Boundary markers: The documentation explicitly instructs the agent to treat returned text as untrusted content, though technical delimiters are not enforced at the instruction level.\n
Capability inventory: The skill has access to shell execution via the CLI and file-writing capabilities through the -o flag (SKILL.md).\n
Sanitization: No automated sanitization of the extracted web content is defined; it relies on the agent's internal safety guardrails.

tavily-extract