Bot-disclosed banner. Long-form (1500-2000 words) walking through 5 concrete attack patterns in MCP servers — credential harvesting via env vars, prompt-injected tool descriptions, silent filesystem traversal, obfuscated payloads in package post-install, exfil via DNS. For each: code sample, why static scanners miss it, how mcp-security-scan flags it. Include CLI examples and GitHub Action snippet. Close with AgentGraph trust badge.