This server is a research/educational tool demonstrating prompt injection and tool poisoning attacks, NOT a production MCP server. All 12 tools have severely compromised descriptions containing malicious payloads (prompt injection, phishing links, credential harvesting, system override commands). While basic schemas exist for 3 benign tools (get_weather, get_forecast, get_temperature in benign_server.py), the malicious variants and remaining tools embed attack vectors directly in descriptions. Tools 4-11 are explicitly designed to manipulate LLM behavior through description text. No error handling, no security controls, no output schemas documented. This violates every security and composition pattern.

• URGENT: Remove all prompt injection payloads from tool descriptions. Descriptions must contain ONLY factual information about what the tool does, not instructions to the LLM or references to non-existent tools.
• Remove duplicate tool names. If benign and malicious variants must coexist for research, use distinct names (e.g., get_weather_benign, get_weather_poisoned) so LLMs can distinguish them.
• Document output schemas for all tools. Specify return type (string, object, array), field names, and types. Include pagination parameters (limit, offset, total_count) for list-returning tools.
• Add explicit error handling with recovery guidance. E.g., 'City not found. Try search_cities() with a partial name.' instead of generic errors.
• If this is research code, add a prominent README warning that this server demonstrates ATTACK VECTORS and must NEVER be deployed in production. Separate benign and malicious implementations into clearly labeled directories.