The Agent Skills Directory

[Indirect Prompt Injection] (LOW): The skill enables agents to fetch and process data from arbitrary external sources, which may contain malicious instructions designed to influence the agent's behavior. Evidence: 1. Ingestion points: Data enters the agent's context via tool outputs from valyu_search and valyu_contents (e.g., in references/integrations/anthropic.md and google.md). 2. Boundary markers: The implementation examples and prompt guides do not consistently demonstrate the use of delimiters or 'ignore' instructions for search results. 3. Capability inventory: Agents configured with these tools typically possess capabilities for synthesis and multi-step tool calling, which can be manipulated by injected instructions. 4. Sanitization: While the system converts HTML to markdown, there is no verification of natural language filtering for adversarial prompts.
[No Code] (SAFE): The analyzed skill consists entirely of Markdown files containing documentation, integration guides, and code snippets; no executable scripts, binaries, or configuration files were provided for analysis.

valyu-best-practices