HumanCraft MCP Server is maintained by AI Build Infra. We take security, prompt integrity, and computational safety seriously.

This Model Context Protocol (MCP) server operates via Stdio transport and interfaces directly with LLM agents like Google Antigravity. As such, our threat model specifically guards against:

1. ReDoS (Regular Expression Denial of Service):

   • Untrusted text (headings, user-submitted HTML) cannot trigger catastrophic backtracking.
   • All pattern matchers operate on length-bounded inputs (max 300 characters per heading, max 100KB per HTML document).
   • Regex patterns are designed with deterministic, linear time complexity $O(n)$.

2. URI & SSRF Injection:

   • The audit_eeat_entity_graph tool parses external sameAs entity identifiers.
   • Strict protocol whitelisting is enforced: only https:// URIs are permitted.
   • Dangerous schemes such as javascript:, data:, file:, ftp:, and unencoded scripts are rejected.

3. XSS & HTML/Attribute Injection:

   • The build_comparison_matrix tool renders HTML <table> elements and Schema.org JSON-LD scripts.
   • All user inputs are sanitized and HTML-entity-escaped (&, <, >, ", ') before being inserted into DOM nodes or attributes.

4. Token Resource Exhaustion & Context Flooding:

   • Hard bounds on tool inputs and array lengths prevent memory exhaustion and runaway token spend.
   • Tool outputs are dense and concise, preserving the agent's context window.

If you discover a security vulnerability in this project:

1. Do NOT open a public GitHub issue.
2. Send an email directly to security@aibuildinfra.com or visit https://aibuildinfra.com/security.
3. Include:
   • Description of the vulnerability
   • Steps to reproduce or proof of concept (PoC)
   • Potential impact
4. We acknowledge receipt within 24 hours and will coordinate a patched release prior to any public disclosure.