The Agent Skills Directory

[SAFE]: The skill is a defensive security guide. While it contains strings associated with prompt injection (e.g., 'Ignore previous instructions'), these are exclusively used within educational examples, code snippets for regex-based scanners, and descriptions of the OWASP LLM Top 10 vulnerabilities.
[DATA_EXFILTRATION]: The skill provides guidance on preventing data exfiltration, specifically addressing SSRF via tool arguments and secret exposure in environment variables. It includes examples of link-local IP deny-lists (e.g., 169.254.169.254) to prevent metadata service probing.
[CREDENTIALS_UNSAFE]: The skill promotes secure credential handling by recommending the use of Cloud Secret Managers (GCP, AWS, HashiCorp Vault) and Pydantic's SecretStr to mask secrets in logs and traces, explicitly advising against storing production secrets in .env files.
[DYNAMIC_EXECUTION]: The documentation warns against using shell tools in production and provides a safe pattern for parameterized, sandboxed command execution as an alternative.
[INDIRECT_PROMPT_INJECTION]: The skill specifically addresses vulnerabilities in RAG chains where untrusted documents might contain instructions, providing the recommended XML-tag boundary and canary token solution.

langchain-security-basics