The Agent Skills Directory

[DATA_EXFILTRATION]: The skill communicates with an external server at mcp.deepwiki.com to retrieve documentation. User-provided repository names and queries are transmitted to this service as part of its core functionality.
[PROMPT_INJECTION]: The SKILL.md and SAFETY.md files contain significant instructional content intended to modify the agent's behavior. This includes direct commands to the agent (e.g., "Never construct URLs", "Confirm with the user before making more than 5 requests") and few-shot examples in SAFETY.md that demonstrate how the agent should refuse certain user requests.
[COMMAND_EXECUTION]: The skill's primary interface is a local Node.js script (./scripts/deepwiki.js). The agent is required to execute this script via the shell, which involves passing arguments such as repository names and search queries into a subprocess.
[INDIRECT_PROMPT_INJECTION]: This category identifies an attack surface where malicious instructions could be hidden in the data being processed.
Ingestion points: Documentation content and wiki structures are fetched from external GitHub repositories via the deepwiki.js script and printed to standard output for the agent to read.
Boundary markers: No explicit boundary markers or delimiters are used in the script's output to separate the untrusted repository content from the agent's instruction context.
Capability inventory: The agent can execute shell commands (running the Node.js script) and has network connectivity through the script wrapper.
Sanitization: There is no evidence of sanitization, filtering, or instruction-detection logic applied to the content retrieved from GitHub before it is presented to the agent.

deepwiki-hardened