The Agent Skills Directory

[COMMAND_EXECUTION]: The skill documents how to launch MCP servers as child processes using standard tools like npx, node, python, or uvx (described in targets/stdio.md). This is the standard operational model for MCP stdio transport.
[PROMPT_INJECTION]: The skill includes an extensive catalog of prompt injection patterns (e.g., in catalog.json under content-injection and cross-resource-leakage) used to test the resilience of target servers against adversarial instructions. These are legitimate test cases for security assessment.
[EXTERNAL_DOWNLOADS]: The skill references external URLs such as wuzzi.net and webhook.site (in catalog.json under ssrf and content-injection). These are well-known security research and testing tools used for verifying Server-Side Request Forgery and indirect prompt injection.
[DATA_EXFILTRATION]: The skill contains patterns for testing data exposure (e.g., reading /etc/passwd, AWS/GCP metadata endpoints). These are explicitly used as fail-criteria and test payloads within the catalog.json to identify vulnerabilities in the target server, not to exfiltrate data from the user's own environment.
[CREDENTIALS_UNSAFE]: The skill contains logic to scan for environment variables and configuration files (like .env, mcp.json) in the local repository to auto-detect server settings. It includes explicit instructions to the agent to NEVER ask for secret values and only extract environment variable names (documented in SKILL.md section 3).

opfor-mcp-setup