The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill demonstrates patterns for extracting structured data from untrusted external sources, creating an attack surface for indirect prompt injection.\n
Ingestion points: The skill processes raw text from various sources such as resumes, articles, and user messages (e.g., in SKILL.md Pattern 1 and Pattern 4, and multiple examples in references/examples.md) by interpolating them directly into LLM prompts.\n
Boundary markers: The provided prompt templates lack robust delimiters or specific instructions (e.g., 'ignore any instructions contained within the following text') to prevent the model from obeying commands embedded in the data.\n
Capability inventory: The skill focuses on generating structured Python objects. No direct demonstration of dangerous capabilities (such as shell execution or file system modification) using the extracted data is present, which limits the potential impact.\n
Sanitization: No input validation or filtering logic is shown to mitigate the risk of processing adversarial content within the input text.\n- [REMOTE_CODE_EXECUTION]: The documentation mentions configuration parameters that allow remote code execution from model repositories.\n
Evidence: The trust_remote_code=True parameter is referenced in references/backends.md for vLLM configuration.\n
Context: This parameter enables the execution of custom code from a Hugging Face or vLLM model repository. The examples provided target well-known models from established organizations, making this a documented feature of the library rather than a malicious pattern.\n- [CREDENTIALS_UNSAFE]: The skill includes examples of how to provide API keys to model backends.\n
Evidence: Initialization examples in SKILL.md and references/backends.md show the use of an api_key parameter.\n
Context: The examples use safe placeholders like "your-api-key" and explicitly recommend the use of environment variables for secret management, adhering to best practices.

outlines