Refusal training for LLM agents against disguised MCP attacks · HackerLangs